Publications

Please visit the PRADA Lab publication page for the full list of publication of our lab.

(★ my Master/PhD/Intern/Visiting students)

Large Model Papers
[Arxiv] Dialectical Alignment: Resolving the Tension of 3H and Security Threats of LLMs [Link] Abstract▼
With the rise of large language models (LLMs), ensuring they embody the principles of being helpful, honest, and harmless (3H), known as Human Alignment, becomes crucial. While existing alignment methods like RLHF, DPO, etc., effectively fine-tune LLMs to match preferences in the preference dataset, they often lead LLMs to highly receptive human input and external evidence, even when this information is poisoned. This leads to a tendency for LLMs to be \textit{Adaptive Chameleons} when external evidence conflicts with their parametric memory. This exacerbates the risk of LLM being attacked by external poisoned data, which poses a significant security risk to LLM system applications such as Retrieval-augmented generation (RAG). To address the challenge, we propose a novel framework: Dialectical Alignment (DA), which (1) utilizes AI feedback to identify optimal strategies for LLMs to navigate inter-context conflicts and context-memory conflicts with different external evidence in context window (i.e., different ratios of poisoned factual contexts); (2) constructs the SFT dataset as well as the preference dataset based on the AI feedback and strategies above; (3) uses the above datasets for LLM alignment to defense poisoned context attack while preserving the effectiveness of in-context knowledge editing. Our experiments show that the dialectical alignment model improves poisoned data attack defense by 20% and does not require any additional prompt engineering or prior declaration of "you may be attacked" to the LLMs' context window.

Shu Yang, Jiayuan Su, Han Jiang, Mengdi Li, Keyuan Cheng, Muhammad Asif Ali, Lijie Hu, Di Wang.
Under review


[Arxiv] PROMPT-SAW: Leveraging Relation-Aware Graphs for Textual Prompt Compression [Link] Abstract▼
Large language models (LLMs) have shown exceptional abilities for multiple different natural language processing tasks. While prompting is a crucial tool for LLM inference, we observe that there is a significant cost associated with exceedingly lengthy prompts. Existing attempts to compress lengthy prompts lead to sub-standard results in terms of readabil- ity/interpretation of the compressed prompt, with a detrimental impact on prompt utility. To address this, we propose PROMPT-SAW: PROMPT compresSion via Relation AWare graphs, an effective strategy for prompt compression over task-agnostic and task-aware prompts. PROMPT-SAW uses the prompt’s textual information to build a graph, later extracts key information elements in the graph to come up with the compressed prompt. We also propose GSM8K-AUG, i.e., an extended version of the existing GSM8k benchmark for task-agnostic prompts in order to provide a comprehensive evaluation platform. Experimental evaluation using benchmark data sets shows that prompts compressed by PROMPT-SAW are not only better in terms of readability, but they also outperform the best-performing baseline models by up to 14.3% and 13.7% respectively for task-aware and task-agnostic settings while compressing the original prompt text by 33.0% and 56.7%.

Muhammad Asif Ali, Zhengping Li, Shu Yang, Keyuan Cheng, Yang Cao, Tianhao Huang, Lijie Hu, Lu Yu, Di Wang.
Under review


[Arxiv] Multi-hop Question Answering under Temporal Knowledge Editing [Link] Abstract▼
Multi-hop question answering (MQA) under knowledge editing (KE) has garnered significant attention in the era of large language models. However, existing models for MQA under KE exhibit poor performance when dealing with questions containing explicit temporal contexts. To address this limitation, we propose a novel framework, namely TEMPoral knowLEdge augmented Multi-hop Question Answering (TEMPLE-MQA). Unlike previous methods, TEMPLE-MQA first constructs a time-aware graph (TAG) to store edit knowledge in a structured manner. Then, through our proposed inference path, structural retrieval, and joint reasoning stages, TEMPLE-MQA effectively discerns temporal contexts within the question query. Experiments on benchmark datasets demonstrate that TEMPLE-MQA significantly outperforms baseline models. Additionally, we contribute a new dataset, namely TKEMQA, which serves as the inaugural benchmark tailored specifically for MQA with temporal scopes.

Keyuan Cheng, Gang Lin, Haoyang Fei, Yuxuan Zhai, Lu Yu, Muhammad Asif Ali, Lijie Hu, Di Wang.
Under review


[Arxiv] Privacy-Preserving Low-Rank Adaptation for Latent Diffusion Models [Link] Abstract▼
Low-rank adaptation (LoRA) is an efficient strategy for adapting latent diffusion models (LDMs) on a training dataset to generate specific objects by minimizing the adaptation loss. However, adapted LDMs via LoRA are vulnerable to membership inference (MI) attacks that can judge whether a particular data point belongs to private training datasets, thus facing severe risks of privacy leakage. To defend against MI attacks, we make the first effort to propose a straightforward solution: privacy-preserving LoRA (PrivateLoRA). PrivateLoRA is formulated as a min-max optimization problem where a proxy attack model is trained by maximizing its MI gain while the LDM is adapted by minimizing the sum of the adaptation loss and the proxy attack model's MI gain. However, we empirically disclose that PrivateLoRA has the issue of unstable optimization due to the large fluctuation of the gradient scale which impedes adaptation. To mitigate this issue, we propose Stable PrivateLoRA that adapts the LDM by minimizing the ratio of the adaptation loss to the MI gain, which implicitly rescales the gradient and thus stabilizes the optimization. Our comprehensive empirical results corroborate that adapted LDMs via Stable PrivateLoRA can effectively defend against MI attacks while generating high-quality images.

Zihao Luo, Xilie Xu, Feng Liu, Yun Sing Koh, Di Wang, and Jingfeng Zhang
Under review


[Arxiv] MoRAL: MoE Augmented LoRA for LLMs’ Lifelong Learning [Link] Abstract▼
Adapting large language models (LLMs) to new domains/tasks and enabling them to be effi- cient lifelong learners is a pivotal challenge. In this paper, we propose MoRAL, i.e., Mixture- of-Experts augmented Low Rank Adaptation for Lifelong learning. MoRAL combines the multi-tasking abilities of MoE with the fine- tuning abilities of LoRA for effective life-long learning of LLMs. In contrast to the con- ventional approaches that use factual triplets as inputs MoRAL relies on simple question- answer pairs, which is a more practical and effective strategy for robust and efficient learn- ing. Owing to new data settings, we introduce a new evaluation benchmark namely: Life Long Learning of LLM (5L-bench) encompassing a newly curated dataset of question-answer pairs, and a set of evaluation metrics for rigorous evaluation of MoRAL in open-book and closed- book settings. Experimental evaluation shows (i) LLMs learn fast in open-book settings with up to 30.15% improvement in "RA" for Phi-2- 2.7B compared to closed-book (for models fine- tuned with MoRAL); (ii) MoRAL shows higher performance improvement for models with a greater number of parameters; (iii) MoRAL is robust to catastrophic forgetting offering better knowledge retention compared to baselines.

Shu Yang, Muhammad Asif Ali, Cheng-Long Wang, Lijie Hu, and Di Wang
Under review


[Arxiv] MONAL: Model Autophagy Analysis for Modeling Human-AI Interactions [Link] Abstract▼
The increasing significance of large models and their multi-modal variants in societal information processing has ignited debates on social safety and ethics. However, there exists a paucity of comprehensive analysis for (i) the interactions between human and artificial intelligence systems, and (ii) understanding and addressing the associated limitations. To bridge this gap, we propose Model AutOphagy ANALysis (MONAL) for self-consumption explanation. MONAL employs two distinct autophagous loops (referred to as “self-consumption loops”) to elucidate the suppression of human-generated information in the exchange between human and AI systems. Through comprehensive experiments on diverse datasets, we evaluate the capacities of generated models as both creators and disseminators of information. Our key findings reveal (i) A progressive prevalence of model-generated synthetic information over time within training datasets compared to human-generated information; (ii) The discernible tendency of large models, when acting as information transmitters across multiple iterations, to selectively modify or prioritize specific contents; and (iii) The potential for a reduction in the diversity of socially or human-generated information, leading to bottlenecks in the performance enhancement of large models and confining them to local optima.

Shu Yang, Muhammad Asif Ali, Lu Yu, Lijie Hu, and Di Wang
Under review


[Arxiv] Towards Personalized AI: Early-stopping Low-Rank Adaptation of Foundation Models [Link] Abstract▼
Foundation models, such as Latent Diffusion Models and Generative Pre-trained Transformers, trained on broad data have shown impressive results in various downstream applications. Fine-tuning a pre-trained foundation model is an affordable way to customize it on small and personalized data. However, the non-AI experts often struggle with the hyperparameter configurations and sometimes encounter the overfitting issue without even realizing it. To mitigate this issue, we introduce a new monitoring metric (CS-Fluctuation) to facilitate early stopping the fine-tuning process. Specifically, we leverage Low-Rank Adaptation (LoRA) to fit the small scale of the personalized data while monitoring the cosine similarity of the parameter changes between the LoRA branch and its corresponding layer. When the changes become steady, we observe the onset of overfitting issue which becomes increasingly severe as fine-tuning progresses. Empirically, we leverage various types of personalized data to conduct customization experiments on both vision and language foundation models, which corroborates the effectiveness of CS-Fluctuation in early stopping the LoRA fine-tuning.

Zihao Luo, Di Wang, Yun Sing Koh, and Jingfeng Zhang
Under review


[Arxiv] Fair Text-to-Image Diffusion Models via Fair Mapping [Link] Abstract▼
In this paper, we address the limitations of existing text-to-image diffusion models in generating demographically fair results when given human-related descriptions. These models often struggle to disentangle the target language context from sociocultural biases, resulting in biased image generation. To overcome this challenge, we propose Fair Mapping, a general, model-agnostic, and lightweight approach that modifies a pre-trained text-to-image model by controlling the prompt to achieve fair image generation. One key advantage of our approach is its high efficiency. The training process only requires updating a small number of parameters in an additional linear mapping network. This not only reduces the computational cost but also accelerates the optimization process. We first demonstrate the issue of bias in generated results caused by language in text-guided diffusion models. By developing a mapping network that projects language embeddings into a debiased space, we enable the generation of relatively balanced demographic results based on the keyword specified in the prompt. With comprehensive experiments on face image generation, we show that our method significantly improves image generation performance when prompted with descriptions related to human faces. By effectively addressing the issue of bias, we produce more fair and diverse image outputs. Our work contributes to the field of text-to-image generation by enhancing the ability to generate images that accurately reflect the intended demographic characteristics specified in the text.

Jia Li, Lijie Hu, Jingfeng Zhang, Tianhang Zheng, Hua Zhang, and Di Wang
Under review


[Arxiv] Fake News Detectors are Biased against Texts Generated by Large Language Models [Link] Abstract▼
The spread of fake news has emerged as a critical challenge, undermining trust and posing threats to society. In the era of Large Language Models (LLMs), the capability to generate believable fake content has intensified these concerns. In this study, we present a novel paradigm to evaluate fake news detectors in scenarios involving both human-written and LLM-generated misinformation. Intriguingly, our findings reveal a significant bias in many existing detectors: they are more prone to flagging LLM-generated content as fake news while often misclassifying human-written fake news as genuine. This unexpected bias appears to arise from distinct linguistic patterns inherent to LLM outputs. To address this, we introduce a mitigation strategy that leverages adversarial training with LLM-paraphrased genuine news. The resulting model yielded marked improvements in detection accuracy for both human and LLM-generated news. To further catalyze research in this domain, we release two comprehensive datasets, \texttt{GossipCop++} and \texttt{PolitiFact++}, thus amalgamating human-validated articles with LLM-generated fake and real news.

Jinyan Su^*, Terry Yue Zhuo^*, Jonibek Mansurov, Di Wang, and Preslav Nakov
Under review


[ICLR] An LLM can Fool Itself: A Prompt-Based Adversarial Attack [Link] Abstract▼
The wide-ranging applications of large language models (LLMs), especially in safety-critical domains, necessitate the proper evaluation of the LLM's adversarial robustness. This paper proposes an efficient tool to audit the LLM's adversarial robustness via a prompt-based adversarial attack (PromptAttack). PromptAttack converts adversarial textual attacks into an attack prompt that can cause the victim LLM to output the adversarial sample to fool itself. The attack prompt is composed of three important components: (1) original input (OI) including the original sample and its ground-truth label, (2) attack objective (AO) illustrating a task description of generating a new sample that can fool itself without changing the semantic meaning, and (3) attack guidance (AG) containing the perturbation instructions to guide the LLM on how to complete the task by perturbing the original sample at character, word, and sentence levels, respectively. Besides, we use a fidelity filter to ensure that PromptAttack maintains the original semantic meanings of the adversarial examples. Further, we enhance the attack power of PromptAttack by ensembling adversarial examples at different perturbation levels. Comprehensive empirical results using Llama2 and GPT-3.5 validate that PromptAttack consistently yields a much higher attack success rate compared to AdvGLUE and AdvGLUE++. Interesting findings include that a simple emoji can easily mislead GPT-3.5 to make wrong predictions.

Xilie Xu, Keyi Kong, Ning Liu, Lizhen Cui, Di Wang, Jingfeng Zhang, and Mohan Kankanhalli
The 12th International Conference on Learning Representations (ICLR 2024)


[EMNLP] DetectLLM: Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Text [Link] Abstract▼
With the rapid progress of Large language models (LLMs) and the huge amount of text they generated, it becomes impractical to manually distinguish whether a text is machine-generated. Given the growing use of LLMs in social media and education, it prompts us to develop methods to detect machine-generated text, preventing malicious use such as plagiarism, misinformation, and propaganda. In this paper, we introduce two novel zero-shot methods for detecting machine-generated text by leveraging the log rank information. One is called DetectLLM-LRR, which is fast and efficient, and the other is called DetectLLM-NPR, which is more accurate, but slower due to the need for perturbations. Our experiments on three datasets and seven language models show that our proposed methods improve over the state of the art by 3.9 and 1.75 AUROC points absolute. Moreover, DetectLLM-NPR needs fewer perturbations than previous work to achieve the same level of performance, which makes it more practical for real-world use. We also investigate the efficiency–performance trade-off based on users' preference on these two measures and provide intuition for using them in practice effectively.

Jinyan Su, Terry Yue Zhuo, Di Wang, and Preslav Nakov
Findings of The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP Findings)

2024

Conference Papers

1. [IEEE S&P] Preserving Node-level Privacy in Graph Neural Networks [Link] Abstract▼

   Differential privacy (DP) has seen immense applications in learning on tabular, image, and sequential data where instance-level privacy is concerned. In learning on graphs, contrastingly, works on node-level privacy are highly sparse. Challenges arise as existing DP protocols hardly apply to the message-passing mechanism in Graph Neural Networks (GNNs). In this study, we propose a solution that specifically addresses the issue of node-level privacy. Our protocol consists of two main components: 1) a sampling routine called HeterPoisson, which employs a specialized node sampling strategy and a series of tailored operations to generate a batch of sub-graphs with desired properties, and 2) a randomization routine that utilizes symmetric multivariate Laplace (SML) noise instead of the commonly used Gaussian noise. Our privacy accounting shows this particular combination provides a non-trivial privacy guarantee. In addition, our protocol enables GNN learning with good performance, as demonstrated by experiments on five real-world datasets; compared with existing baselines, our method shows significant advantages, especially in the high privacy regime. Experimentally, we also 1) perform membership inference attacks against our protocol and 2) apply privacy audit techniques to confirm our protocol's privacy integrity. In the sequel, we present a study on a seemingly appealing approach (USENIX'23) that protects node-level privacy via differentially private node/instance embeddings. Unfortunately, such work has fundamental privacy flaws, which are identified through a thorough case study. More importantly, we prove an impossibility result of achieving both (strong) privacy and (acceptable) utility through private instance embedding. The implication is that such an approach has intrinsic utility barriers when enforcing differential privacy.

   
   Zihang Xiang^★, Tianhao Wang, Di Wang
   The 45th IEEE Symposium on Security and Privacy (IEEE S&P 2024).
   


2. [VLDB] Privacy Amplification via Shuffling: Unified, Simplified, and Tightened Abstract▼

   The shuffle model of differential privacy provides promising privacy-utility balances in decentralized, privacy-preserving data analysis. However, the current analyses of privacy amplification via shuffling lack both tightness and generality. To address this issue, we propose the \emph{variation-ratio reduction} as a comprehensive framework for privacy amplification in both single-message and multi-message shuffle protocols. It leverages two new parameterizations: the total variation bounds of local messages and the probability ratio bounds of blanket messages, to determine indistinguishability levels. Our theoretical results demonstrate that our framework provides tighter bounds, especially for local randomizers with extremal probability design, where our bounds are exactly tight. Additionally, variation-ratio reduction complements parallel composition in the shuffle model, yielding enhanced privacy accounting for popular sampling-based randomizers employed in statistical queries (e.g., range queries, marginal queries, and frequent itemset mining). Empirical findings demonstrate that our numerical amplification bounds surpass existing ones, conserving up to $30\%$ of the budget for single-message protocols, $75\%$ for multi-message ones, and a striking $75\%$-$95\%$ for parallel composition. Our bounds also result in a remarkably efficient $\tilde{O}(n)$ algorithm that numerically amplifies privacy in less than $10$ seconds for $n=10^8$ users.

   
   Shaowei Wang, Yun Peng, Jin Li, Zikai Wen, Zhipeng Li, Shiyu Yu, Di Wang, and Wei Yang
   International Conference on Very Large Data Bases (VLDB 2024)
   


3. [VLDB] Communication Efficient and Provable Federated Unlearning Abstract▼

   We investigate the problem of federated unlearning, which aims to erase the influence of requested clients or data samples on the global model trained via federated learning (FL). This problem arises from the increasing demand for the right to be forgotten as well as various privacy threats in FL. We propose a novel exact federated unlearning framework that satisfies two crucial desiderata: \textit{communication efficiency} and \textit{exact unlearning provability}. To the best of our knowledge, our work is the first to address both aspects in a unified manner. We first introduce a rigorous definition of \textit{exact} federated unlearning, which ensures that the unlearned model is statistically indistinguishable from the one trained without the removed data. We then identify the key property to enable rapid exact federated unlearning: total variation (TV) stability, which captures the sensitivity of the model parameters to small changes in the dataset. Based on this insight, we design a TV-stable FL algorithm called \texttt{FATS}, which adapts the classical \texttt{\underline{F}ed\underline{A}vg} algorithm for \underline{T}V \underline{S}tability and incorporates local SGD with periodic averaging to reduce the communication round. We also devise efficient unlearning algorithms for \texttt{FATS} under two scenarios: client-level and sample-level unlearning. We provide theoretical guarantees for our learning and unlearning algorithms, showing that they achieve exact federated unlearning with reasonable convergence rates for both the original and unlearned models. We empirically evaluate our framework on 6 benchmark datasets, and demonstrate its superiority over state-of-the-art baselines in terms of accuracy, communication cost, computation cost, and unlearning effectiveness.

   
   Youming Tao^★*, Chenglong Wang^★*, Miao Pan, Dongxiao Yu, Xiuzhen Cheng, and Di Wang
   International Conference on Very Large Data Bases (VLDB 2024)
   


4. [ICLR] Faithful Vision-Language Interpretation via Concept Bottleneck Models [Link] Abstract▼

   The demand for transparency in healthcare and finance has led to interpretable machine learning (IML) models, notably the concept bottleneck models (CBMs), valued for its potential in performance and insights into deep neural networks. However, CBM's reliance on manually annotated data poses challenges. Label-free CBM has emerged to address this, but they remain unstable, affecting their faithfulness as explanatory tools. To address this inherent instability issue, we introduce a formal definition for an alternative concept called the Faithful Vision-Language Concept (FVLC) models. We present a methodology for constructing an FVLC that satisfies four critical properties. Our extensive experimentation, conducted on four benchmark datasets using Label-free CBM model architectures, demonstrates that our FVLC outperforms other baselines in terms of stability against input and concept set perturbations. Our approach incurs minimal accuracy degradation compared to the vanilla CBM, making it a promising solution for reliable and faithful model interpretation.

   
   Songning Lai^★*, Lijie Hu^★*, Junxiao Wang, Laure Berti-Equille, and Di Wang
   The 12th International Conference on Learning Representations (ICLR 2024)
   


5. [ICLR] Improved Analysis of Sparse Linear Regression in Local Differential Privacy Model [Link] Abstract▼

   In this paper, we revisit the problem of sparse linear regression in the local differential privacy (LDP) model. Existing research in the non-interactive and sequentially local models has focused on obtaining the lower bounds for the case where the underlying parameter is $1$-sparse, and extending such bounds to the more general $k$-sparse case has proven to be challenging. Moreover, it is unclear whether efficient non-interactive LDP (NLDP) algorithms exist. To address these issues, we first consider the problem in the $\epsilon$ non-interactive LDP model and provide a lower bound of $\Omega(\frac{\sqrt{dk\log d}}{\sqrt{n}\epsilon})$ on the $\ell_2$-norm estimation error for sub-Gaussian data, where $n$ is the sample size and $d$ is the dimension of the space. We propose an innovative NLDP algorithm, the very first of its kind for the problem. As a remarkable outcome, this algorithm also yields a novel and highly efficient estimator as a valuable by-product. Our algorithm achieves an upper bound of $\tilde{O}({\frac{d\sqrt{k}}{\sqrt{n}\epsilon}})$ for the estimation error when the data is sub-Gaussian, which can be further improved by a factor of $O(\sqrt{d})$ if the server has additional public but unlabeled data. For the sequentially interactive LDP model, we show a similar lower bound of $\Omega({\frac{\sqrt{dk}}{\sqrt{n}\epsilon}})$. As for the upper bound, we rectify a previous method and show that it is possible to achieve a bound of $\tilde{O}(\frac{k\sqrt{d}}{\sqrt{n}\epsilon})$. Our findings reveal fundamental differences between the non-private case, central DP model, and local DP model in the sparse linear regression problem.

   
   Liyang Zhu^★*, Meng Ding^★*, Vaneet Aggarwal, Jinhui Xu, and Di Wang
   The 12th International Conference on Learning Representations (ICLR 2024)
   


6. [ICLR] Theoretical Analysis of Robust Overfitting for Wide DNNs: An NTK Approach [Link] Abstract▼

   Adversarial training (AT) is a canonical method for enhancing the robustness of deep neural networks (DNNs). However, recent studies empirically demonstrated that it suffers from robust overfitting, i.e., a long time AT can be detrimental to the robustness of DNNs. This paper presents a theoretical explanation of robust overfitting for DNNs. Specifically, we non-trivially extend the neural tangent kernel (NTK) theory to AT and prove that an adversarially trained wide DNN can be well approximated by a linearized DNN. Moreover, for squared loss, closed-form AT dynamics for the linearized DNN can be derived, which reveals a new AT degeneration phenomenon: a long-term AT will result in a wide DNN degenerates to that obtained without AT and thus cause robust overfitting. Based on our theoretical results, we further design a method namely Adv-NTK, the first AT algorithm for infinite-width DNNs. Experiments on real-world datasets show that Adv-NTK can help infinite-width DNNs enhance comparable robustness to that of their finite-width counterparts, which in turn justifies our theoretical findings.

   
   Shaopeng Fu^★ and Di Wang
   The 12th International Conference on Learning Representations (ICLR 2024)
   


7. [ICLR] An LLM can Fool Itself: A Prompt-Based Adversarial Attack [Link] Abstract▼

   The wide-ranging applications of large language models (LLMs), especially in safety-critical domains, necessitate the proper evaluation of the LLM's adversarial robustness. This paper proposes an efficient tool to audit the LLM's adversarial robustness via a prompt-based adversarial attack (PromptAttack). PromptAttack converts adversarial textual attacks into an attack prompt that can cause the victim LLM to output the adversarial sample to fool itself. The attack prompt is composed of three important components: (1) original input (OI) including the original sample and its ground-truth label, (2) attack objective (AO) illustrating a task description of generating a new sample that can fool itself without changing the semantic meaning, and (3) attack guidance (AG) containing the perturbation instructions to guide the LLM on how to complete the task by perturbing the original sample at character, word, and sentence levels, respectively. Besides, we use a fidelity filter to ensure that PromptAttack maintains the original semantic meanings of the adversarial examples. Further, we enhance the attack power of PromptAttack by ensembling adversarial examples at different perturbation levels. Comprehensive empirical results using Llama2 and GPT-3.5 validate that PromptAttack consistently yields a much higher attack success rate compared to AdvGLUE and AdvGLUE++. Interesting findings include that a simple emoji can easily mislead GPT-3.5 to make wrong predictions.

   
   Xilie Xu, Keyi Kong, Ning Liu, Lizhen Cui, Di Wang, Jingfeng Zhang, and Mohan Kankanhalli
   The 12th International Conference on Learning Representations (ICLR 2024)
   


8. [EACL] Antonym vs Synonym Distinction using InterlaCed Encoder NETworks (ICE-NET) [Link] Abstract▼

   Antonyms vs synonyms distinction is a core challenge in lexico-semantic analysis and automated lexical resource construction. These pairs share a similar distributional context which makes it harder to distinguish them. Leading research in this regard attempts to capture the properties of the relation pairs, i.e., symmetry, transitivity, and trans-transitivity. However, the inability of existing research to appropriately model the relation-specific properties limits their end performance. In this paper, we propose InterlaCed Encoder NETworks (i.e., ICE-NET) for antonym vs synonym distinction, that aim to capture and model the relation-specific properties of the antonyms and synonyms pairs in order to perform the classification task in a performance-enhanced manner. Experimental evaluation using the benchmark datasets shows that ICE-NET outperforms the existing research by a relative score of upto 1.8% in F1-measure.

   
   Muhammad Asif Ali^★, Yan Hu^★, Jianbin Qin, and Di Wang
   The 18th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2024 Findings)
   


9. [EACL] Differentially Private Natural Language Models: Recent Advances and Future Directions [Link] Abstract▼

   Recent developments in deep learning have led to great success in various natural language processing (NLP) tasks. However, these applications may involve data that contain sensitive information. Therefore, how to achieve good performance while also protect privacy of sensitive data is a crucial challenge in NLP. To preserve privacy, Differential Privacy (DP), which can prevent reconstruction attacks and protect against potential side knowledge, is becoming a de facto technique for private data analysis. In recent years, NLP in DP models (DP-NLP) has been studied from different perspectives, which deserves a comprehensive review. In this paper, we provide the first systematic review of recent advances on DP deep learning models in NLP. In particular, we first discuss some differences and additional challenges of DP-NLP compared with the standard DP deep learning. Then we investigate some existing work on DP-NLP and present its recent developments from two aspects: gradient perturbation based methods and embedding vector perturbation based methods. We also discuss some challenges and future directions of this topic.

   
   Lijie Hu^★, Ivan Habernal, Lei Shen, and Di Wang
   The 18th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2024 Findings)
   



Journal Papers

10. [JMLR] Faster Rates of Private Stochastic Convex Optimization [Link] Abstract▼

    In this paper, we revisit the problem of Differentially Private Stochastic Convex Optimization (DP-SCO) and provide excess population risks for some special classes of functions that are faster than the previous results of general convex and strongly convex functions. In the first part of the paper, we study the case where the population risk function satisfies the Tysbakov Noise Condition (TNC) with some parameter $\theta>1$. Specifically, we first show that under some mild assumptions on the loss functions, there is an algorithm whose output could achieve an upper bound of $\tilde{O}((\frac{1}{\sqrt{n}}+\frac{d}{n\epsilon})^\frac{\theta}{\theta-1}) $ and $\tilde{O}((\frac{1}{\sqrt{n}}+\frac{\sqrt{d\log(1/\delta)}}{n\epsilon})^\frac{\theta}{\theta-1})$ for $\epsilon$-DP and $(\epsilon, \delta)$-DP, respectively when $\theta\geq 2$, here $n$ is the sample size and $d$ is the dimension of the space. Then we address the inefficiency issue, improve the upper bounds by $\text{Poly}(\log n)$ factors and extend to the case where $\theta\geq \bar{\theta}>1$ for some known $\bar{\theta}$. Next we show that the excess population risk of population functions satisfying TNC with parameter $\theta\geq 2$ is always lower bounded by $\Omega((\frac{d}{n\epsilon})^\frac{\theta}{\theta-1}) $ and $\Omega((\frac{\sqrt{d\log(1/\delta)}}{n\epsilon})^\frac{\theta}{\theta-1})$ for $\epsilon$-DP and $(\epsilon, \delta)$-DP, respectively, which matches our upper bounds. In the second part, we focus on a special case where the population risk function is strongly convex. Unlike the previous studies, here we assume the loss function is {\em non-negative} and {\em the optimal value of population risk is sufficiently small}. With these additional assumptions, we propose a new method whose output could achieve an upper bound of $O(\frac{d\log(1/\delta)}{n^2\epsilon^2}+\frac{1}{n^{\tau}})$ and $O(\frac{d^2}{n^2\epsilon^2}+\frac{1}{n^{\tau}})$ for any $\tau> 1$ in $(\epsilon,\delta)$-DP and $\epsilon$-DP model respectively if the sample size $n$ is sufficiently large. These results circumvent their corresponding lower bounds in \cite{feldman2020private} for general strongly convex functions. Finally, we conduct experiments of our new methods on real world data. Experimental results also provide new insights into established theories.

    
    Jinyan Su^★, Lijie Hu^★, and Di Wang
    Journal of Machine Learning Research
    


11. [TMC] Private Over-the-Air Federated Learning at Band-Limited Edge [Link] Abstract▼

    We investigate over-the-air federated learning (OTA-FL) that exploits over-the-air computing (AirComp) to integrate communication and computation seamlessly for FL. Privacy presents a serious obstacle for OTA-FL, as it can be compromised by maliciously manipulating channel state information (CSI). Moreover, the limited band at edge hinders OTA-FL from training large-scale models. It remains open how to enable a multitude of devices with constrained resources and sensitive data to collaboratively train a global model at band-limited edge. To tackle this, we design a novel algorithm PROBE building upon a lightweight over-the-air gradients aggregation rule PB-O-GAR. Specifically, PB-O-GAR combines a random sparsification-like dimension reduction with Gaussian perturbation to provide rigorous privacy and band-adapted communication. It elaborately calibrates the transmission signal according to devices’ perceived CSI for heterogeneous power constraints accommodation and CSI attack resilience. We show that by utilizing the common randomness, which deviates from the conventional FL, random sparsification-like dimension reduction can augment privacy in addition to the intrinsic privacy amplification effect of AirComp. We establish near-optimal convergence rates and explicit trade-offs among privacy, communication and utility for PROBE. Finally, extensive experiments on benchmark datasets are conducted to validate our theoretical findings and showcase the superiority of PROBE in realistic settings

    
    Youming Tao, Shuzhen Chen, Congwei Zhang, Di Wang, Dongxiao Yu, Xiuzhen Cheng, and Falko Dressler.
    Revision, IEEE Transactions on Mobile Computing.
    


12. [TKDD] Fair Single Index Model [Link] Abstract▼

    Single-index models (SIMs) have been widely used in various applications due to their simplicity and interpretability. However, despite the potential for SIMs to result in discriminatory outcomes based on sensitive attributes like gender, race, or ethnicity, the issue of fairness has not been thoroughly examined in recent studies on the topic. This paper aims to address these fairness concerns by proposing methods for building fair SIMs. Specifically, based on the definition of equal opportunity, we first provide a fairness definition for SIM. Next, we develop a unified fair SIM model and propose an efficient method to solve the fair SIM. Theoretically, we also show that our output is consistent in fairness. Finally, we conduct comprehensive experimental studies over 7 benchmark datasets and demonstrate that our fair SIM outperforms the other 8 baseline methods.

    
    Yidong Wang^★*, Meng Ding^★*, Jinhui Xu and Di Wang
    Revision, ACM Transactions on Knowledge Discovery from Data
    


13. [TMLR] Persistent Local Homology in Graph Learning [Link] Abstract▼

    In this study, we introduce Persistent Local Homology (PLH) for graphs, a novel method that synergizes persistent homology with local homology to analyze graph structures. We begin by mathematically formalizing PLH, defining it as the application of persistent homology to annular local subgraphs. This foundation paves the way for the development of a computational pipeline, specifically tailored for PLH, which we explore in various graph learning contexts. Despite its utility, a complexity analysis reveals potential computational bottlenecks in PLH application. To address this, we propose Reduced PLH (rPLH), an efficient variant designed to significantly lower computational complexity. Experimental evaluations with rPLH demonstrate its capability to retain the effectiveness of the original PLH while substantially reducing computational demands. The practical utility of PLH and rPLH is further corroborated through comprehensive experiments on both synthetic and real-world datasets, highlighting their broad applicability and potential in diverse analytical scenarios.

    
    Minghua Wang, Yan Hu, Ziyun Huang, Di Wang, and Jinhui Xu
    Transactions on Machine Learning Research
    


14. [TCBB] A Network Enhancement Method to Identify Spurious Drug-Drug Interactions [Link] Abstract▼

    As medical safety and drug regulation gain heightened attention, the detection of spurious drug-drug interactions (DDI) has become key in healthcare. Although current research using graph neural networks (GNNs) to predict DDI has shown impressive results, it often fails to account for false DDI in the constructed DDI networks. Such inaccuracies caused by data errors, false alarms, or incorrect drug details can skew the network’s structure and hinder the accuracy of GNN-based predictions. To tackle this challenge, we propose ANSM, a network-enhancement method specifically designed to identify and attenuate spurious links between drugs for ensuring the accuracy of DDI networks. ANSM integrates three key components: the feature extractor, the network optimizer, and the discriminative classifier. The feature extractor captures local structural features from drug node pairs, while the network optimizer leverages network information to improve feature extraction and reduce the impact of spurious DDI links. The discriminative classifier then identifies potential spurious links. Experimental results demonstrate that ANSM outperforms state-of-the-art methods in identifying spurious DDI.

    
    Huan Wang^★, Ziwen Cui, Yinguang Yang, Baijing Wang, Lida Zhu, Wen Zhang, Pan Zhou, and Di Wang
    IEEE/ACM Transactions on Computational Biology and Bioinformatics
    


15. [TBD] A Multi-classification Division-aggregation Framework for Fake News Detection [Link] Abstract▼

    Nowadays, as human activities are shifting to social media, fake news detection has been a crucial problem. Existing methods ignore the classification difference in online news and cannot take full advantage of multi-classification knowledges. For example, when coping with a post “A mouse is frightened by a cat,” a model that learns “computer” knowledge tends to misunder- stand “mouse” and give a fake label, but a model that learns “ani- mal” knowledge tends to give a true label. Therefore, this research proposes a multi-classification division-aggregation framework to detect fake news, named CKA, which innovatively learns classifi- cation knowledges during training stages and aggregates them during prediction stages. It consists of three main components: a news characterizer, an ensemble coordinator, and a truth predic- tor. The news characterizer is responsible for extracting news fea- tures and obtaining news classifications. Cooperating with the news characterizer, the ensemble coordinator generates classifica- tion-specifical models for the maximum reservation of classifica- tion knowledges during the training stage, where each classifica- tion-specifical model maximizes the detection performance of fake news on corresponding news classifications. Further, to aggregate the classification knowledges during the prediction stage, the truth predictor uses the truth discovery technology to aggregate the pre- dictions from different classification-specifical models based on re- liability evaluation of classification-specifical models. Extensive experiments prove that our proposed CKA outperforms state-of- the-art baselines in fake news detection.

    
    Wen Zhang, Haitao Fu, Lionel Z. Wang, Huan Wang^★, Zhiguo Gong, Pan Zhou, and Di Wang
    IEEE Transactions on Big Data
    


16. [NL] Near-perfect Coverage Manifold Estimation in Cellular Networks via conditional GAN [Link] Abstract▼

    This paper presents a conditional generative adver- sarial network (cGAN) that translates base station location (BSL) information of any Region-of-Interest (RoI) to location-dependent coverage probability values within a subset of that region, called the region-of-evaluation (RoE). We train our network utilizing the BSL data of India, the USA, Germany, and Brazil. In comparison to the state-of-the-art convolutional neural networks (CNNs), our model improves the prediction error (L1 difference between the coverage manifold generated by the network under consideration and that generated via simulation) by two orders of magnitude. Moreover, the cGAN-generated coverage manifolds appear to be almost visually indistinguishable from the ground truth.

    
    Washim Uddin Mondal, Veni Goyal, Goutam Das, Satish V. Ukkusuri, Di Wang, Mohamed-Slim Alouini, and Vaneet Aggarwal
    IEEE Networking Letters
    



2023

Conference Papers

17. [USENIX] Inductive Graph Unlearning [Link] [Code] Abstract▼

    As a way to implement the "right to be forgotten" in machine learning, \textit{machine unlearning} aims to completely remove the contributions and information of the samples to be deleted from a trained model without affecting the contributions of other samples. Recently, many frameworks for machine unlearning have been proposed, and most of them focus on image and text data. To extend machine unlearning to graph data, \textit{GraphEraser} has been proposed. However, a critical issue is that \textit{GraphEraser} is specifically designed for the transductive graph setting, where the graph is static and attributes and edges of test nodes are visible during training. {\color{blue}It is unsuitable for the inductive setting, where the graph could be dynamic and the test graph information is invisible in advance. Such inductive capability is essential for production machine learning systems with evolving graphs like social media and transaction networks.} To fill this gap, we propose the \underline{{\bf G}}\underline{{\bf U}}ided Guided Inductive Graph Unlearning framework (GUIDE). GUIDE consists of three components: guided graph partitioning with fairness and balance, efficient subgraph repair, and similarity-based aggregation. Empirically, we evaluate our method on several inductive benchmarks and evolving transaction graphs. Experimental results demonstrate the efficiency and practicality of GUIDE for inductive graph unlearning. In particular, GUIDE achieves faster implementation compared to \textit{GraphEraser} and outperforms the existing state-of-the-art methods on inductive graph learning tasks.

    
    Cheng-Long Wang^★, Mengdi Huai, and Di Wang
    The 32nd USENIX Security Symposium (USENIX 2023)
    


18. [IEEE S&P] A Theory to Instruct Differentially-Private Learning via Clipping Bias Reduction [Link] [Code] Abstract▼

    We study the bias introduced in Differentially-Private Stochastic Gradient Descent (DP-SGD) with clipped or normalized per-sample gradient. As one of the most popular but artificial operations to ensure bounded sensitivity, gradient clipping enables composite privacy analysis of many iterative optimization methods without additional assumptions on either learning models or input data. Despite its wide applicability, gradient clipping presents theoretical challenges in systematically instructing improvement of privacy or utility. In general, without an assumption on globally-bounded gradient, classic convergence analyses do not apply to clipped gradient descent. Further, given limited understanding of the utility loss, many existing improvements upon DP-SGD are heuristic, especially in the applications of private deep learning. In this paper, we provide meaningful theoretical analysis validated by thorough empirical results of DP-SGD. We point out that the bias caused by gradient clipping is highly underestimated in previous works. For generic non-convex optimization via DP-SGD, we show one key factor contributing to the bias is the sampling noise of stochastic gradient to be clipped. Accordingly, we use the developed theory to build a series of improvements for sampling noise reduction from various perspectives. From an optimization angle, we study variance reduction techniques and propose inner-outer momentum. At the learning model (neural network) level, we propose several tricks to enhance network internal normalization and BatchClipping to carefully clip the gradient of a batch of samples. For the data preprocessing, we provide theoretical justification of recently proposed improvements via data normalization and (self-)augmentation. Putting these systematic improvements together, private deep learning via DP-SGD can be significantly strengthened in many tasks. For example, in computer vision applications, with (\epsilon=8, \delta=10^{-5}) DP guarantee, we successfully train Resnet20 on CIFAR10 with accuracy 74.5%; for natural language processing, with (\epsilon=4, \delta=10^{-5}), we successfully train recurrent neural network on IMDb data with accuracy 77.5%.

    
    Hanshen Xiao^*, Zihang Xiang^*★, Di Wang, and Srini Devadas (* equal contribution)
    The 44th IEEE Symposium on Security and Privacy (IEEE S&P 2023)
    


19. [SIGMOD] On Practical Differentially Private and Byzantine-resilient Federated Learning [Link] [Code] Abstract▼

    Privacy and Byzantine resilience are two indispensable requirements for a federated learning (FL) system. Although there have been extensive studies on privacy and Byzantine security in their own track, solutions that consider both remain sparse. This is due to difficulties in reconciling privacy-preserving and Byzantine-resilient algorithms. In this work, we propose a solution to such a two-fold issue. We use our version of differentially private stochastic gradient descent (DP-SGD) algorithm to preserve privacy and then apply our Byzantine-resilient algorithms. We note that while existing works follow this general approach, an in-depth analysis on the interplay between DP and Byzantine resilience has been ignored, leading to unsatisfactory performance. Specifically, for the random noise introduced by DP, previous works strive to reduce its seemingly detrimental impact on the Byzantine aggregation. In contrast, we leverage the random noise to construct a first-stage aggregation that effectively rejects many existing Byzantine attacks. Moreover, based on another property of our DP variant, we form a second-stage aggregation which provides a final sound filtering. Our protocol follows the principle of co-designing both DP and Byzantine resilience. We provide both theoretical proof and empirical experiments to show our protocol is effective: retaining high accuracy while preserving the DP guarantee and Byzantine resilience. Compared with the previous work, our protocol 1) achieves significantly higher accuracy even in a high privacy regime; 2) works well even when up to 90% distributive workers are Byzantine.

    
    Zihang Xiang^★, Tianhao Wang , Wanyu Lin, and Di Wang
    International Conference on Management of Data (SIGMOD 2023)
    


20. [NeurIPS] On Private and Robust Bandits [Link] Abstract▼

    We study private and robust multi-armed bandits (MABs), where the agent receives Huber's contaminated heavy-tailed rewards and meanwhile needs to ensure differential privacy. We consider both the finite $k$-th raw moment and the finite $k$-th central moment settings for heavy-tailed rewards distributions with $k\ge 2$. We first present its minimax lower bound, characterizing the information-theoretic limit of regret with respect to privacy budget, contamination level, and heavy-tailedness. Then, we propose a meta-algorithm that builds on a private and robust mean estimation sub-routine \texttt{PRM} that essentially relies on reward truncation and the Laplace mechanism. For the above two different heavy-tailed settings, we give corresponding schemes of \texttt{PRM}, which enable us to achieve nearly-optimal regrets. Moreover, our two proposed truncation-based or histogram-based \texttt{PRM} schemes achieve the optimal trade-off between estimation accuracy, privacy and robustness. Finally, we support our theoretical results and show the effectiveness of our algorithms with experimental studies.

    
    Yulian Wu^★*, Xingyu Zhou^*, Youming Tao and Di Wang
    2023 Conference on Neural Information Processing Systems (NeurIPS 2023)
    


21. [ICML] Differentially Private Episodic Reinforcement Learning with Heavy-tailed Rewards [Link] Abstract▼

    In this paper we study the problem of (finite horizon tabular) Markov decision processes (MDPs) with heavy-tailed rewards under the constraint of differential privacy (DP). Compared with the previous studies for private reinforcement learning that typically assume rewards are sampled from some bounded or sub-Gaussian distributions to ensure DP, we consider the setting where reward distributions have only finite (1+v)-th moments with some $v\in (0, 1]$. By resorting to robust mean estimators for rewards, we first propose two frameworks for heavy-tailed MDPs, i.e., one is for value iteration and another is for policy optimization. Under each framework, we consider both joint differential privacy (JDP) and local differential privacy (LDP) models. Based on our frameworks, we provide regret upper bounds for both JDP and LDP cases, and show that the moment of distributions and privacy budget have significant impact on regrets. Finally, we establish a lower bound of regret minimization for heavy-tailed MDPs in JDP model by reducing it to the instance-independent lower bound of heavy-tailed multi-armed bandits in DP model. We also show the lower bound for the problem in LDP by adopting some private minimax methods. Our results reveal that there are fundamental differences between the problem of private RL with sub-Gaussian and that with heavy-tailed rewards.

    
    Yulian Wu^★, Xingyu Zhou, Sayak Ray Chowdhury and Di Wang
    The 40th International Conference on Machine Learning (ICML 2023)
    


22. [AISTATS] Privacy-preserving Sparse Generalized Eigenvalue Problem [Link] Abstract▼

    In this paper we study the (sparse) Generalized Eigenvalue Problem (GEP), which arises in a number of modern statistical learning models, such as principal component analysis (PCA), canonical correlation analysis (CCA), Fisher's discriminant analysis (FDA) and sliced inverse regression (SIR). Existing techniques for GEP all fail to consider the protection of sensitive information in the training set. Models learned by such methods can implicitly memorize the details of sensitive information. To address this issue, we provide the first study on GEP in the differential privacy (DP) model under both deterministic and stochastic settings. In the low dimensional case, we provide a $\rho$-Concentrated DP (CDP) method namely DP-Rayleigh Flow and show if the initial vector is close enough to the optimal vector, its output has an $\ell_2$-norm estimation error of $\tilde{O}(\frac{d}{n}+\frac{d}{n^2\rho})$ (under some mild assumptions), where $d$ is the dimension and $n$ is the sample size. Next, we discuss how to find such a initial parameter privately. In the high dimensional sparse case where $d\gg n$, we propose the DP-Truncated Rayleigh Flow method whose output could achieve an error of $\tilde{O}(\frac{s\log d}{n}+\frac{s\log d}{n^2\rho})$ for various statistical models, where $s$ is the sparsity of the underlying parameter. Moreover, we show that these errors in the stochastic setting are optimal up to a factor of $\text{Poly}(\log n)$ by providing the lower bounds of PCA and SIR under statistical setting and in the CDP model. Finally, to give a separation between $\epsilon$-DP and $\rho$-CDP for GEP, we also provide the lower bound $\Omega(\frac{d}{n}+\frac{d^2}{n^2\epsilon^2})$ and $\Omega(\frac{s\log d}{n}+\frac{s^2\log^2 d}{n^2\epsilon^2})$ of private minimax risk for PCA, under the statistical setting and $\epsilon$-DP model, in low and high dimensional sparse case respectively. Experiments on both synthetic and real-world data also support our theoretical analysis.

    
    Lijie Hu^*★, Zihang Xiang^*★, Jiabin Liu, and Di Wang (* equal contribution)
    The 26th International Conference on Artificial Intelligence and Statistics (AISTATS 2023)
    


23. [AAAI] SEAT: Stable and Explainable Attention [Link] Abstract▼

    Currently, attention mechanism becomes a standard fixture in most state-of-the-art natural language processing (NLP) models, not only due to outstanding performance it could gain, but also due to plausible innate explanation for the behaviors of neural architectures it provides, which is notoriously difficult to analyze. However, recent studies show that attention is unstable against randomness and perturbations during training or testing, such as random seeds and slight perturbation of embedding vectors, which impedes it from becoming a faithful explanation tool. Thus, a natural question is whether we can find some substitute of the current attention which is more stable and could keep the most important characteristics on explanation and prediction of attention. In this paper, to resolve the problem, we provide a first rigorous definition of such alternate namely SEAT (Stable and Explainable Attention). Specifically, a SEAT should has the following three properties: (1) Its prediction distribution is enforced to be close to the distribution based on the vanilla attention; (2) Its top-$k$ indices have large overlaps with those of the vanilla attention; (3) It is robust w.r.t perturbations, i.e., any slight perturbation on SEAT will not change the prediction distribution too much, which implicitly indicates that it is stable to randomness and perturbations. Moreover we propose a method to get a SEAT, which could be considered as an ad hoc modification for the canonical attention. Finally, through intensive experiments on various datasets, we compare our SEAT with other baseline methods using RNN, BiLSTM and BERT architectures via six different evaluation metrics for model interpretation, stability and accuracy. Results show that SEAT is more stable against different perturbations and randomness while also keeps the explainability of attention, which indicates it is a more faithful explanation. Moreover, compared with vanilla attention, there is almost no utility (accuracy) degradation for SEAT.

    
    Lijie Hu^*★, Yixin Liu ^*, Ninghao Liu , Mengdi Huai, Lichao Sun, and Di Wang (* equal contribution)
    The 37th AAAI Conference on Artificial Intelligence (AAAI 2023)
    Selected as an Oral paper
    


24. [UAI] Differentially Private Stochastic Convex Optimization in (Non)-Euclidean Space Revisited [Link] Abstract▼

    In this paper, we revisit the problem of Differentially Private Stochastic Convex Optimization (DP-SCO) in Euclidean and general $\ell_p^d$ spaces. Specifically, we focus on three settings that are still far from well understood: (1) DP-SCO over a constrained and bounded (convex) set in Euclidean space; (2) unconstrained DP-SCO in $\ell_p^d$ space; (3) DP-SCO with heavy-tailed data over a constrained and bounded set in $\ell_p^d$ space. For problem (1), for both convex and strongly convex loss functions, we propose methods whose outputs could achieve (expected) excess population risks that are only dependent on the Gaussian width of the constraint set rather than the dimension of the space. Moreover, we also show the bound for strongly convex functions is optimal up to a logarithmic factor. For problems (2) and (3), we propose several novel algorithms and provide the first theoretical results for both cases.

    
    Jinyan Su^★, Changhong Zhao and Di Wang
    The 39th Conference on Uncertainty in Artificial Intelligence (UAI 2023)
    
    
25. [MobiSys] High-Speed Wireless Communications Inspired Energy Efficient Federated Learning over Mobile Devices [Link] Abstract▼

    Energy efficiency is essential for federated learning (FL) over mobile devices and its potential prosperous applications. Different from existing communication efficient FL research efforts, which regard communication energy consumption as the bottleneck, we have observed that with ever increasing wireless transmission speed (e.g., Wi-Fi 5 or 5G), the energy consumption of wireless communications for model updates in FL is significantly reduced and sometimes is smaller than that of local on-device training. Motivated by such observations, in this paper, we propose a high-speed wireless communications inspired energy efficient federated learning over mobile devices (EEFL), whose goal is to reduce the overall energy consumption (computing + communication). In particular, we design a novel energy-aware adaptive local update policy for mobile devices by jointly considering FL performance and energy saving of high-speed wireless transmissions. Furthermore, given the device's local update policy in each FL global round, we advance the dynamic voltage and frequency scaling (DVFS) strategy to minimize local training's energy consumption by keeping GPU and CPU working at appropriate frequencies without triggering thermal throttling. Extensive experimental results with various learning models, datasets, and wireless transmission environments demonstrate the proposed EEFL's superiority over the peer designs in terms of energy efficiency.

    
    Rui Chen, Qiyu Wan, Xinyue Zhang, Xiaoqi Qin, Di Wang, Xin Fu, and Miao Pan
    The 21st ACM International Conference on Mobile Systems, Applications, and Services (MobiSys 2023)
    
    
26. [EMNLP] GRI: Graph-based Relative Isomorphism of Word Embedding Spaces [Link] Abstract▼

    Automated construction of bi-lingual dictionaries using monolingual embedding spaces is a core challenge in machine translation. The end performance of these dictionaries relies on the geometric similarity of individual spaces, i.e., their degree of isomorphism. Existing attempts aimed at controlling the relative isomorphism of different spaces fail to incorporate the impact of lexically different but semantically related words in the training objective. To address this, we propose GRI that combines the distributional training objectives with attentive graph convolutions to unanimously consider the impact of lexical variations of semantically similar words required to define/compute the relative isomorphism of multiple spaces. Exper imental evaluation shows that GRI outperforms the existing research by improving the average P@1 by a relative score of upto 63.6%.

    
    Muhammad Asif Ali^★, Yan Hu^★, Jianbin Qin, and Di Wang
    Findings of The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP Findings)
    
    
27. [EMNLP] DetectLLM: Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Text [Link] Abstract▼

    With the rapid progress of Large language models (LLMs) and the huge amount of text they generated, it becomes impractical to manually distinguish whether a text is machine-generated. Given the growing use of LLMs in social media and education, it prompts us to develop methods to detect machine-generated text, preventing malicious use such as plagiarism, misinformation, and propaganda. In this paper, we introduce two novel zero-shot methods for detecting machine-generated text by leveraging the log rank information. One is called DetectLLM-LRR, which is fast and efficient, and the other is called DetectLLM-NPR, which is more accurate, but slower due to the need for perturbations. Our experiments on three datasets and seven language models show that our proposed methods improve over the state of the art by 3.9 and 1.75 AUROC points absolute. Moreover, DetectLLM-NPR needs fewer perturbations than previous work to achieve the same level of performance, which makes it more practical for real-world use. We also investigate the efficiency–performance trade-off based on users' preference on these two measures and provide intuition for using them in practice effectively.

    
    Jinyan Su, Terry Yue Zhuo, Di Wang, and Preslav Nakov
    Findings of The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP Findings)
    
    
28. [ECAI] Finite Sample Guarantees of Differentially Private Expectation Maximization Algorithm [Link] Abstract▼

    (Gradient) Expectation Maximization (EM) is a widely used algorithm for estimating the maximum likelihood of mixture models or incomplete data problems. A major challenge facing this popular technique is how to effectively preserve the privacy of sensitive data. Previous research on this problem has already lead to the discovery of some Differentially Private (DP) algorithms for (Gradient) EM. However, unlike in the non-private case, existing techniques are not yet able to provide finite sample statistical guarantees. To address this issue, we propose in this paper the first DP version of Gradient EM algorithm with statistical guarantees. Specifically, we first propose a new mechanism for privately estimating the mean of a heavy-tailed distribution, which significantly improves a previous result in \cite{wangicml2020}, and it could be extended to the local DP model, which has not been studied before. Next, we apply our general framework to three canonical models: Gaussian Mixture Model (GMM), Mixture of Regressions Model (MRM) and Linear Regression with Missing Covariates (RMC). Specifically, for GMM in the DP model, our estimation error is near optimal in some cases. For the other two models, we provide the first result on finite sample statistical guarantees. Our theory is supported by thorough numerical experiments on both real-world data and synthetic data.

    
    Di Wang^*, Jiahao Ding^*, Lijie Hu, Zejun Xie, Miao Pan, and Jinhui Xu
    The 26th European Conference on Artificial Intelligence (ECAI 2023)
    
    
29. [ArabicNLP] GARI: Graph Attention for Relative Isomorphism of Arabic Word Embeddings. [Link] Abstract▼

    Bilingual Lexical Induction (BLI) is a core challenge in NLP, it relies on the relative isomorphism of individual embedding spaces. Existing attempts aimed at controlling the relative isomorphism of different embedding spaces fail to incorporate the impact of semantically related words in the model training objective. To address this, we propose GARI that combines the distributional training objectives with multiple isomorphism losses guided by the graph attention network. GARI considers the impact of semantical variations of words in order to define the relative isomorphism of the embedding spaces. Experimental evaluation using the Arabic language data set shows that GARI outperforms the existing research by improving the average P@1 by a relative score of up to 40.95% and 76.80% for in-domain and domain mismatch settings respectively.

    
    Muhammad Asif Ali^★, Maha Alshmrani^★, Jianbin Qin, Yan Hu^★, and Di Wang
    The First Arabic Natural Language Processing Conference (ArabicNLP 2023)
    
    

    Journal Papers

30. [JMLR] Generalized Linear Models in Non-interactive Local Differential Privacy with Public Data [Link] Abstract▼

    In this paper, we study the problem of estimating smooth Generalized Linear Models (GLMs) in the Non-interactive Local Differential Privacy (NLDP) model. Different from its classical setting, our model allows the server to access some additional public but unlabeled data. In the first part of the paper we focus on GLMs. Specifically, we first consider the case where each data record is i.i.d. sampled from a zero-mean multivariate Gaussian distribution. Motivated by the Stein's lemma, we present an $(\epsilon, \delta)$-NLDP algorithm for GLMs. Moreover, the sample complexity of the public and private data for the algorithm to achieve an $\ell_2$-norm estimation error of $\alpha$ (with high probability) is ${O}(p \alpha^{-2})$ and $\tilde{O}(p^3\alpha^{-2}\epsilon^{-2})$ respectively, where $p$ is the dimension of the feature vector. This is a significant improvement over the previously known exponential or quasi-polynomial in $\alpha^{-1}$, or exponential in $p$ sample complexities of GLMs with no public data. Then we consider a more general setting where each data record is i.i.d. sampled from some sub-Gaussian distribution with bounded $\ell_1$-norm. Based on a variant of Stein's lemma, we propose an $(\epsilon, \delta)$-NLDP algorithm for GLMs (under some mild assumptions). The sample complexity of the public and private data for the algorithm to achieve an $\ell_\infty$-norm estimation error of $\alpha$ is ${O}(p^2\alpha^{-2})$ and $\tilde{O}(p^2\alpha^{-2}\epsilon^{-2})$ respectively, if $\alpha$ is not too small ({\em i.e.,} $\alpha\geq \Omega(\frac{1}{\sqrt{p}})$). In the second part of the paper, we extend our idea to the non-linear regression problem and show similar results for it under the multivariate Gaussian and sub-Gaussian cases. Finally, we demonstrate the effectiveness of our algorithms through experiments on both synthetic and real world datasets. To our best knowledge, this is the first paper showing the existence of efficient and effective algorithms for GLMs and non-linear regression in the NLDP model with public unlabeled data.
    

    
    Di Wang^*, Lijie Hu^*★, Huanyu Zhang, Marco Gaboardi, and Jinhui Xu (* equal contribution)
    Journal of Machine Learning Research, Volume 24, 132 (2023), Pages 1-57


31. [TIT] Quantizing Heavy-tailed Data in Statistical Estimation:(Near) Minimax Rates, Covariate Quantization, and Uniform Recovery [Link] Abstract▼

    This paper studies the quantization of heavy-tailed data in some fundamental statistical estimation problems, where the underlying distributions have bounded moments of some order. We propose to truncate and properly dither the data prior to a uniform quantization. Our major standpoint is that (near) minimax rates of estimation error are achievable merely from the quantized data produced by the proposed scheme. In particular, concrete results are worked out for covariance estimation, compressed sensing, and matrix completion, all agreeing that the quantization only slightly worsens the multiplicative factor. Besides, we study compressed sensing where both covariate (i.e., sensing vector) and response are quantized. Under covariate quantization, although our recovery program is non-convex because the covariance matrix estimator lacks positive semi-definiteness, all local minimizers are proved to enjoy near optimal error bound. Moreover, by the concentration inequality of product process and covering argument, we establish near minimax uniform recovery guarantee for quantized compressed sensing with heavy-tailed noise.

    
    Junren Chen^★, Michael Kwok Po NG, and Di Wang
    IEEE Transactions on Information Theory
    


32. [TIT] High Dimensional Statistical Estimation under Uniformly Dithered One-bit Quantization [Link] Abstract▼

    In this paper, we consider several fundamental machine learning and (sparse) statistical estimation problems from i.i.d samples of a sub-Gaussian or heavy-tailed distribution. Instead of having the full knowledge of the samples, we focus on the scenario where each entry of these the samples is quantized to one or two bits, which plays a critical role in signal processing or machine learning applications. Specifically, we give the first study on the problems of (sparse) covariance matrix estimation, sparse linear regression and low-rank matrix completion. For covariance matrix estimation, compared with the previous results, we extend to the heavy-tailed case where the underlying distribution has bounded fourth order moments. Moreover, we extend to the high dimensional sparse (and heavy-tailed) cases by providing several new estimators which consist of four steps: Truncating, Dithering, Quantizing and Thresholding. For sparse linear model and low-rank matrix completion, we propose new quadratic objective functions for one-bit quantized samples based on our previous estimators. For all of our estimators we also derive estimation errors. Especially, in the sub-Gaussian case, all of our estimators could achieve near optimal rates in the unquantized cases. Finally, extensive experiments on synthetic and real-world data have been carried out to support our theories.

    
    Junren Chen^★, Cheng-Long Wang^★, Michael Kwok Po NG, and Di Wang
    IEEE Transactions on Information Theory, Volume 69, 8 (2023), Pages 5151-5187
    


33. [Science Advances] PPML-Omics: a Privacy-Preserving federated Machine Learning System Protects Patients’ Privacy from Omic Data [Link] Abstract▼

    Modern machine learning models towards various tasks with omic data analysis give rise to threats of privacy leakage of patients involved in those datasets. Despite the advances in different privacy technologies, existing methods tend to introduce too much noise, which hampers model accuracy and usefulness. Here, we built a secure and privacy-preserving machine learning (PPML) system by combining federated learning (FL), differential privacy (DP) and shuffling mechanism. We applied this system to analyze data from three sequencing technologies, and addressed the privacy concern in three major tasks of omic data, namely cancer classification with bulk RNA-seq, clustering with single-cell RNA-seq, and the integration of spatial gene expression and tumour morphology with spatial transcriptomics, under three representative deep learning models. We also examined privacy breaches in depth through privacy attack experiments and demonstrated that our PPML-Omics system could protect patients’ privacy. In each of these applications, PPML-Omics was able to outperform state-of-the-art systems under the same level of privacy guarantee, demonstrating the versatility of the system in simultaneously balancing the privacy-preserving capability and utility in omic data analysis. Furthermore, we gave the theoretical proof of the privacy-preserving capability of PPML-Omics, suggesting the first mathematically guaranteed model with robust and generalizable empirical performance.

    
    Juexiao Zhou^*, Siyuan Chen^*, Yulian Wu^*★, Haoyang Li, Bin Zhang, Longxi Zhou, Yan Hu, Zihang Xiang, Zhongxiao Li, Ningning Chen, Wenkai Han, Di Wang, and Xin Gao (* equal contribution)
    Science Advances
    


34. [TKDE] Nearly Optimal Rates of Privacy-preserving Sparse Generalized Eigenvalue Problem [Link] Abstract▼

    In this paper we study the (sparse) Generalized Eigenvalue Problem (GEP), which arises in a number of modern statistical learning models, such as principal component analysis (PCA), canonical correlation analysis (CCA), Fisher's discriminant analysis (FDA) and sliced inverse regression (SIR). We provide the first study on GEP in the differential privacy (DP) model under both deterministic and stochastic settings. In the low dimensional case, we provide a $\rho$-Concentrated DP (CDP) method namely DP-Rayleigh Flow and show if the initial vector is close enough to the optimal vector, its output has an $\ell_2$-norm estimation error of $\tilde{O}(\frac{d}{n}+\frac{d}{n^2\rho})$ (under some mild assumptions), where $d$ is the dimension and $n$ is the sample size. Next, we discuss how to find such an initial parameter privately. In the high dimensional sparse case where $d\gg n$, we propose the DP-Truncated Rayleigh Flow method whose output could achieve an error of $\tilde{O}(\frac{s\log d}{n}+\frac{s\log d}{n^2\rho})$ for various statistical models, where $s$ is the sparsity of the underlying parameter. Moreover, we show that these errors in the stochastic setting are optimal up to a factor of $\text{Poly}(\log n)$ by providing the lower bounds of PCA and SIR under the statistical setting and in the CDP model. Finally, to give a separation between $\epsilon$-DP and $\rho$-CDP for GEP, we also provide the lower bound $\Omega(\frac{d}{n}+\frac{d^2}{n^2\epsilon^2})$ and $\Omega(\frac{s\log d}{n}+\frac{s^2\log^2 d}{n^2\epsilon^2})$ of private minimax risk for PCA, under the statistical setting and $\epsilon$-DP model, in low and high dimensional sparse case respectively. Finally, extensive experiments on both synthetic and real-world data support our previous theoretical analysis.

    
    Lijie Hu^*★, Zihang Xiang^*★, Jiabin Liu, and Di Wang (* equal contribution)
    IEEE Transactions on Knowledge and Data Engineering
    


35. [JCSS] PAC Learning Halfspaces in Non-interactive Local Differential Privacy Model with Public Unlabeled Data Abstract▼

    In this paper, we study the problem of PAC learning halfspaces in the non-interactive local differential privacy model (NLDP). To breach the barrier of exponential sample complexity, previous results studied a relaxed setting where the server has access to some additional public but unlabeled data. We continue in this direction. Specifically, we consider the problem under the standard setting instead of the large margin setting studied before. Under different mild assumptions on the underlying data distribution, we propose two approaches that are based on the Massart noise model and self-supervised learning and show that it is possible to achieve sample complexities that are only linear in the dimension and polynomial in other terms for both private and public data, which significantly improve the previous results. Our methods could also be used for other private PAC learning problems.

    
    Jinyan Su^★, Jinhui Xu, and Di Wang
    Journal of Computer and System Sciences
    


36. [CBM] Personalized and Privacy-preserving Federated Heterogeneous Medical Image Analysis with PPPML-HMI [Link] Abstract▼

    Heterogeneous data is endemic due to the use of diverse models and settings of devices by hospitals in the field of medical imaging. However, there are few open- source frameworks for federated heterogeneous medical image analysis with personalization and privacy protection without the demand to modify the existing model structures or to share any private data. Here, we proposed PPPML-HMI, a novel open- source learning paradigm for personalized and privacy-preserving federated heterogeneous medical image analysis. To our best knowledge, personalization and privacy protection were discussed simultaneously for the first time under the federated scenario by integrating the PerFedAvg algorithm and designing the novel cyclic secure aggregation with the homomorphic encryption algorithm. To show the utility of PPPML- HMI, we applied it to a simulated classification task namely the classification of healthy people and patients from the RAD-ChestCT Dataset, and one real-world segmentation task namely the segmentation of lung infections from COVID-19 CT scans. Meanwhile, we applied the improved deep leakage from gradients to simulate adversarial attacks and showed the strong privacy-preserving capability of PPPML-HMI. By applying PPPML-HMI to both tasks with different neural networks, a varied number of users, and sample sizes, we demonstrated the strong generalizability of PPPML-HMI in privacy-preserving federated learning on heterogeneous medical images.

    
    Juexiao Zhou^*, Longxi Zhou^*, Di Wang, Xiaopeng Xu, Haoyang Li, Yuetan Chu, Wenkai Han, and Xin Gao
    Computers in Biology and Medicine
    


37. [TCS] Gradient Complexity and Non-stationary Views of Differentially Private Empirical Risk Minimization Abstract▼

    In this paper we consider the Differentially Private Empirical Risk Minimization (DP-ERM) problem with either convex or non-convex loss functions. For the cases that DP-ERM with smooth (strongly) convex loss functions with/without (non)-smooth regularization, we propose several methods that achieve (near) optimal expected excess risks ({\em i.e.,} utility bounds) with less gradient complexity compared to the previous methods. For DP-ERM with smooth convex loss functions in high dimensions, we give an algorithm which achieves an upper bound with less gradient complexity than previous ones. For DP-ERM with non-convex loss functions, we consider the problem in both low and high dimensional spaces. For the low dimensional case with non-smooth regularizer, we generalize an existing work by using $\ell_2$ norm of the projected gradient to measure the utility. We also extend the error bound measurement, for the first time, from empirical risk to population risk by using the expected $\ell_2$ norm of the gradient. For the high dimensional case, we first show that by measuring the utility with Frank-Wolfe gap, it is possible to bound the utility by the Gaussian Width of the constraint set, instead of the dimensionality $p$ of the underlying space. We then demonstrate that the advantages of this result can be achieved by measuring the $\ell_2$ norm of the projected gradient. We finally show that the utility of some special non-convex loss functions can be reduced to a level ({\em i.e.,} depending only on $\log p$) similar to that of convex loss functions.

    
    Di Wang, and Jinhui Xu
    Theoretical Computer Science
    



2022

Conference Papers

38. [PODS] High Dimensional Differentially Private Stochastic Optimization with Heavy-tailed Data [Link] Abstract▼

    As one of the most fundamental problems in machine learning, statistics and differential privacy, Differentially Private Stochastic Convex Optimization (DP-SCO) has been extensively studied in recent years. However, most of the previous work can only handle either regular data distribution or irregular data in the low dimensional space case. To better understand the challenges arising from irregular data distribution, in this paper we provide the first study on the problem of DP-SCO with heavy-tailed data in the high dimensional space. In the first part we focus on the problem over some polytope constraint (such as the $\ell_1$-norm ball). We show that if the loss function is smooth and its gradient has bounded second order moment, it is possible to get an error bound of $\tilde{O}(\frac{\log d}{(n\epsilon)^\frac{1}{3}})$ in the $\epsilon$-DP model, where $n$ is the sample size and $d$ is the dimensionality of the underlying space. Next, for LASSO, if the data distribution that has bounded fourth-order moments, we improve the bound to $\tilde{O}(\frac{\log d}{(n\epsilon)^\frac{2}{5}})$ in the $(\epsilon, \delta)$-DP model. In the second part of the paper, we study DP-SCO for sparse learning with heavy-tailed data. We first revisit the sparse linear regression and propose a truncated DP-IHT method whose output could achieve an error of $\tilde{O}(\frac{s^{*2}\log d}{n\epsilon})$, where $s^*$ is the sparsity of the underlying parameter. Then we study a more general problem over the sparsity ({\em i.e.,} $\ell_0$-norm) constraint, and show that it is possible to achieve an error of $\tilde{O}(\frac{s^{*\frac{3}{2}}\log d}{n\epsilon})$, which is also near optimal up to a factor of $\tilde{O}{(\sqrt{s^*})}$, if the loss function is smooth and strongly convex.

    
    Lijie Hu^★, Shuo Ni^★, Hanshen Xiao, and Di Wang
    The 41st ACM Symposium on Principles of Database Systems (PODS 2022)
    Invited to The ACM Transactions on Database Systems special issue on Best of PODS 2022
    ACM CCS 2021 Workshop on Privacy Preserving Machine Learning
    


39. [WINE] Truthful Generalized Linear Models [Link] Abstract▼

    In this paper we study estimating Generalized Linear Models (GLMs) in the case where the agents (individuals) are strategic or self-interested and they concern about their privacy when reporting data. Compared with the classical setting, here we aim to design mechanisms that can both incentivize most agents to truthfully report their data and preserve the privacy of individuals' reports, while their outputs should also close to the underlying parameter. In the first part of the paper, we consider the case where the covariates are sub-Gaussian and the responses are heavy-tailed where they only have the finite fourth moments. First, motivated by the stationary condition of the maximizer of the likelihood function, we derive a novel private and closed form estimator. Based on the estimator, we propose a mechanism which has the following properties via some appropriate design of the computation and payment scheme for several canonical models such as linear regression, logistic regression and Poisson regression: (1) the mechanism is $o(1)$-jointly differentially private (with probability at least $1-o(1)$); (2) it is an $o(\frac{1}{n})$-approximate Bayes Nash equilibrium for a $(1-o(1))$-fraction of agents to truthfully report their data, where $n$ is the number of agents; (3) the output could achieve an error of $o(1)$ to the underlying parameter; (4) it is individually rational for a $(1-o(1))$ fraction of agents in the mechanism ; (5) the payment budget required from the analyst to run the mechanism is $o(1)$. In the second part, we consider the linear regression model under more general setting where both covariates and responses are heavy-tailed and only have finite fourth moments. By using an $\ell_4$-norm shrinkage operator, we propose a private estimator and payment scheme which have similar properties as in the sub-Gaussian case.

    
    Yuan Qiu^★, Jinyan Liu, and Di Wang
    The 18th Conference on Web and Internet Economics (WINE 2022)


40. [ALT] Faster Rates of Private Stochastic Convex Optimization [Link] Abstract▼

    In this paper, we revisit the problem of Differentially Private Stochastic Convex Optimization (DP-SCO) and provide excess population risks for some special classes of functions that are faster than the previous results of general convex and strongly convex functions. In the first part of the paper, we study the case where the population risk function satisfies the Tysbakov Noise Condition (TNC) with some parameter $\theta>1$. Specifically, we first show that under some mild assumptions on the loss functions, there is an algorithm whose output could achieve an upper bound of $\tilde{O}((\frac{1}{\sqrt{n}}+\frac{d}{n\epsilon})^\frac{\theta}{\theta-1}) $ and $\tilde{O}((\frac{1}{\sqrt{n}}+\frac{\sqrt{d\log(1/\delta)}}{n\epsilon})^\frac{\theta}{\theta-1})$ for $\epsilon$-DP and $(\epsilon, \delta)$-DP, respectively when $\theta\geq 2$, here $n$ is the sample size and $d$ is the dimension of the space. Then we address the inefficiency issue, improve the upper bounds by $\text{Poly}(\log n)$ factors and extend to the case where $\theta\geq \bar{\theta}>1$ for some known $\bar{\theta}$. Next we show that the excess population risk of population functions satisfying TNC with parameter $\theta\geq 2$ is always lower bounded by $\Omega((\frac{d}{n\epsilon})^\frac{\theta}{\theta-1}) $ and $\Omega((\frac{\sqrt{d\log(1/\delta)}}{n\epsilon})^\frac{\theta}{\theta-1})$ for $\epsilon$-DP and $(\epsilon, \delta)$-DP, respectively, which matches our upper bounds. In the second part, we focus on a special case where the population risk function is strongly convex. Unlike the previous studies, here we assume the loss function is {\em non-negative} and {\em the optimal value of population risk is sufficiently small}. With these additional assumptions, we propose a new method whose output could achieve an upper bound of $O(\frac{d\log(1/\delta)}{n^2\epsilon^2}+\frac{1}{n^{\tau}})$ and $O(\frac{d^2}{n^2\epsilon^2}+\frac{1}{n^{\tau}})$ for any $\tau> 1$ in $(\epsilon,\delta)$-DP and $\epsilon$-DP model respectively if the sample size $n$ is sufficiently large. These results circumvent their corresponding lower bounds in \cite{feldman2020private} for general strongly convex functions. Finally, we conduct experiments of our new methods on real world data. Experimental results also provide new insights into established theories.

    
    Jinyan Su^★, Lijie Hu^★, and Di Wang
    The 33rd International Conference on Algorithmic Learning Theory (ALT 2022)


41. [AISTATS] Optimal Rates of (Locally) Differentially Private Heavy-tailed Multi-Armed Bandits [Link] Abstract▼

    In this paper we study the problem of stochastic multi-armed bandits (MAB) in the (local) differential privacy (DP/LDP) model. Unlike the previous results which need to assume bounded reward distributions, here we mainly focus on the case the reward distribution of each arm only has $(1+v)$-th moment with some $v\in (0, 1]$. In the first part, we study the problem in the central $\epsilon$-DP model. We first provide a near-optimal result by developing a private and robust Upper Confidence Bound (UCB) algorithm. Then, we improve the result via a private and robust version of the Successive Elimination (SE) algorithm. Finally, we show that the instance-dependent regret bound of our improved algorithm is optimal by showing its lower bound. In the second part of the paper, we study the problem in the $\epsilon$-LDP model. We propose an algorithm which could be seen as locally private and robust version of the SE algorithm, and show it could achieve (near) optimal rates for both instance-dependent and instance-independent regrets. All of the above results can also reveal the differences between the problem of private MAB with bounded rewards and heavy-tailed rewards. To achieve these (near) optimal rates, we develop several new hard instances and private robust estimators as byproducts, which might could be used to other related problems. Finally, experimental results also support our theoretical analysis and show the effectiveness of our algorithms.

    
    Youming Tao^*★, Yulian Wu^*★, Peng Zhao, and Di Wang (* equal contribution)
    The 25th International Conference on Artificial Intelligence and Statistics (AISTATS 2022)
    Selected as an Oral paper (Acceptance Rate: 44/1685=2.6%)
    ACM CCS 2021 Workshop on Privacy Preserving Machine Learning
    ICML 2022 Workshop on Responsible Decision Making in Dynamic Environments (Selected as Contributed Talk)
    


42. [AISTATS] On Facility Location Problem in Local Differential Privacy Model [Link] Abstract▼

    This paper studies the facility location problem, where the given input is combined of (1) a metric space (V, d), (2) facility cost f_i for each facility, We study the uncapacitated facility location (UFL) problem under the constraints imposed by the local differential privacy (LDP). Recently, Gupta et al. (2009) and Esencayi et al. (2019) proposed lower and upper bounds for the UFL problem on the central differential privacy (DP) model where a curator first collects all data before being processed. In this paper, we focus on the LDP model, where we protect a client's participation in the facility location instance. Under the HST metric, we show that there is a non-interactive $\epsilon$-LDP algorithm achieving $O(n^{1/4}/\epsilon^2)$-approximation ratio, where n is the size of the metric. On the negative side, we show a lower bound of $\Omega(n^{1/4}/\sqrt{\epsilon})$ on the approximation ratio for any non-interactive $\epsilon$-LDP algorithm. Thus, our results are tight up to a factor of $\epsilon$. Moreover, unlike previous results, our results generalize for non-uniform facility costs.

    
    [alphabetic order] Vincent Cohen-Addad, Yunus Esencayi, Chenglin Fan, Marco Gaboradi, Shi Li, and Di Wang
    The 25th International Conference on Artificial Intelligence and Statistics (AISTATS 2022)


43. [IJCAI] Private Stochastic Convex Optimization and Sparse Learning with Heavy-tailed Data Revisited [Link] Abstract▼

    In this paper, we revisit the problem of Differentially Private Stochastic Convex Optimization (DP-SCO) with heavy-tailed data, where the gradient of the loss function has bounded moments. Instead of the case where the loss function is Lipschitz or each coordinate of the gradient has bounded second moment studied previously, we consider a relaxed scenario where each coordinate of the gradient only has bounded $(1+v)$-th moment with some $v\in (0, 1]$. Firstly, we start from the one dimensional private mean estimation for heavy-tailed distributions. We propose a novel robust and private mean estimator which is optimal. Based on its idea, we then extend to the general $d$-dimensional space and use it to DP-SCO with general convex and strongly convex loss functions. We also provide lower bounds for these two classes of loss under our setting and show that our upper bounds are optimal up to a factor of $O(\text{Poly}(d))$. To address the high dimensionality issue, we also study DP-SCO with heavy-tailed gradient under some sparsity constraint (DP sparse learning). We propose a new method and show it is also optimal up to a factor of $O(s^*)$, where $s^*$ is the underlying sparsity of the constraint.

    
    Youming Tao^★, Yulian Wu^★, Xiuzhen Cheng, and Di Wang
    The 31st International Joint Conference on Artificial Intelligence (IJCAI-ECAI 2022)


44. [ACML] On PAC Learning Halfspaces in Non-interactive Local Privacy Model with Public Unlabeled Data [Link] Abstract▼

    In this paper, we study the problem of PAC learning halfspaces in the non-interactive local differential privacy model (NLDP). To breach the barrier of exponential sample complexity, previous results studied a relaxed setting where the server has access to some additional public but unlabeled data. We continue along this direction. Specifically, we consider the problem under the standard setting instead of the large margin setting studied before. Under different mild assumptions on the underlying data distribution, we propose two approaches that are based on the Massart noise model and self-supervised learning, and show that it is possible to achieve sample complexities that are only linear in the dimension and polynomial in other terms for both private and public data, which significantly improve the previous results. Our methods might could also be used to other private PAC learning problems.

    
    Jinyan Su^★, Jinhui Xu, and Di Wang.
    The 14th Asian Conference on Machine Learning (ACML 2022)
    Best Paper Award


45. [ISIT] Differentially Private $\ell_1$-norm Linear Regression with Heavy-tailed Data [Link] Abstract▼

    We consider Differentially Private Stochastic Convex Optimization (DP-SCO) with heavy-tailed data. Specifically, we focus on the $\ell_1$ regression in the $\epsilon$-DP model. While most of the previous work focuses on the case where the loss function is Lipschitz, here we only need to assume the variates has bounded moments. Firstly, we study the case where the $\ell_2$ norm of data has bounded second order moment. We propose an algorithm which is based on the exponential mechanism and show that it is possible to achieve an upper bound of $O(\sqrt{\frac{d}{n\epsilon}})$ (with high probability). Next, we relax the assumption to bounded $\theta$-th order moment with some $\theta\in (0, 1)$ and we show that it is possible to achieve an upper bound of $O(({\frac{d}{n\epsilon}})^\frac{\theta-1}{\theta})$. Our algorithms can also be extended to more relaxed cases where only each coordinate of the data has bounded moments.

    
    Di Wang and Jinhui Xu
    2022 IEEE International Symposium on Information Theory (ISIT 2022)



2021

Conference Papers

46. [ALT] Estimating Smooth GLM in Non-interactive Local Differential Privacy Model with Public Unlabeled Data [Link] Abstract▼

    In this paper, we study the problem of estimating smooth Generalized Linear Models (GLM) in the Non-interactive Local Differential Privacy (NLDP) model. Different from its classical setting, our model allows the server to process some additional public but unlabeled data. We first show that there is an $(\epsilon, \delta)$-NLDP algorithm for GLM (under some mild assumptions), if each data record is i.i.d sampled from some sub-Gaussian distribution with bounded $\ell_1$-norm. The sample complexity of both public and private data, for the algorithm to achieve an $\alpha$ estimation error (in $\ell_\infty$-norm), is $\tilde{O}(p^2\alpha^{-2}\epsilon^{-2})$ if $\alpha$ is not too small (i.e., $\alpha\geq \Omega(\frac{1}{\sqrt{p}})$), where $p$ is the dimensionality of the data. This is a significant improvement over the previously known quasi-polynomial (in $\alpha$) or exponential (in $p$) complexity of convex GLM with no public data. We then extend our idea to the non-linear regression problem and show a similar phenomenon for it. Finally, we demonstrate the practicality of our algorithms through experiments on both synthethic and real world datasets. To our best knowledge, this is the first paper showing the existence of efficient and practical algorithms for GLM and non-linear regression in the NLDP model with public unlabeled data.
    

    
    Di Wang^*, Huanyu Zhang^*, Marco Gaboardi and Jinhui Xu (* equal contribution)
    The 32nd International Conference on Algorithmic Learning Theory (ALT 2021)
    NeurIPS 2019 Workshop on Privacy in Machine Learning


47. [IJCAI] Differentially Private Pairwise Learning Revisited [Link] Abstract▼

    Instead of learning with pointwise loss functions, learning with pairwise loss functions (pairwise learning) has received much attention recently as it is more capable of modeling the relative relationship between pairs of samples. However, most of the existing algorithms for pairwise learning fail to take into consideration the privacy issue in their design. To address this issue, previous work studied pairwise learning in the Differential Privacy (DP) model. However, their utilities (population errors) are far from optimal. To address the sub-optimal utility issue, in this paper, we proposed new $(\epsilon, \delta)$ or $\epsilon$-DP algorithms for pairwise learning. Specifically, when the loss functions are Lipschitz, smooth and strongly convex, we show that the output of our algorithm achieves an expected population error of $O(\frac{1}{n}+\frac{d\log\frac{1}{\delta}}{n^2\epsilon^2})$ and $O(\frac{1}{n}+\frac{d^2}{n^2\epsilon^2})$ for $(\epsilon, \delta)$-DP and $\epsilon$-DP, respectively, where $n$ is the sample size and $d$ is the dimension of the underlying space. Moreover, for general convex case, the output of our algorithm achieves an expected population error of $O(\frac{1}{\sqrt{n}}+\frac{\sqrt{d \log\frac{1}{\delta}}}{n\epsilon})$ and $O(\frac{1}{\sqrt{n}}+\frac{d}{n\epsilon})$ for $(\epsilon, \delta)$-DP and $\epsilon$-DP, respectively. It is also notable that these upper bounds are {\bf optimal} ({\em i.e.,} match the lower bounds). We also conduct extensive experiments on real-world datasets to evaluate the proposed algorithms, experimental results support our theoretical analysis and show the priority of our algorithms.
    

    
    Zhiyu Xue^*★, Shaoyang Yang^*★, Mengdi Huai and Di Wang (* equal contribution)
    The 30th International Joint Conference on Artificial Intelligence (IJCAI 2021)



Journal Papers

48. [TIT] On Sparse Linear Regression in the Local Differential Privacy Model [Link] Abstract▼

    In this paper, we study the sparse linear regression problem in the Local Differential Privacy (LDP) model. We first show that polynomial dependency on the dimensionality $p$ of the space is unavoidable for the estimation error in both non-interactive and sequential interactive local models, if the privacy of the whole dataset needs to be preserved. Similar limitations also exist for other types of error measurements and in the relaxed local models. This indicates that differential privacy in high dimensional space is unlikely achievable for the problem. With the understanding of this limitation, we then present two algorithmic results. The first one is a sequential interactive LDP algorithm for the low dimensional sparse case, called Locally Differentially Private Iterative Hard Thresholding (LDP-IHT), which achieves a near optimal upper bound. This algorithm is actually rather general and can be used to solve quite a few other problems, such as (Local) DP-ERM with sparsity constraints and sparse regression with non-linear measurements. The second one is for the restricted (high dimensional) case where only the privacy of the responses (labels) needs to be preserved. For this case, we show that the optimal rate of the error estimation can be made logarithmically depending on $p$ (i.e., $\log p$) in the local model, where an upper bound is obtained by a label-privacy version of LDP-IHT. Experiments on real world and synthetic datasets confirm our theoretical analysis.

    
    Di Wang and Jinhui Xu
    IEEE Transactions on Information Theory, Volume 67, no. 2, Pages 1182-1200, Feb. 2021
    


49. [TCS] Inferring Ground Truth From Crowdsourced Data Under Local Attribute Differential Privacy [Link] Abstract▼

    Recently, the problem of ground truth inference under local differential privacy (LDP) model has been recently studied. However, this problem is still not well understood and even some basic questions have not been solved yet. First, it is still unknown what is the average error of the private estimators to the underlying ground truth. Secondly, we do not known whether we can infer the ability of each user under LDP model and what is the estimation error w.r.t the underlying users ability. Finally, previous work only show that their methods have better performance than the private major voting algorithm through experiments. However, there is still no theoretically result which shows this priority formally or mathematically. In this paper, we partially solve these problems by studying the ground truth inference problem under local attribute differential privacy (LADP) model, and propose a new algorithm called private Dawid-Skene method, which is motivated by the classical Dawid-Skene method. Specifically, we first provide the estimation errors for both ability of users and the ground truth under some assumptions of the problem if the algorithm start with some appropriate initial vector. Moreover, we propose an explicit instance and show that the estimation error of the ground truth achieved by the private major voting algorithm is always greater than the error achieved by our method. To our best knowledge, this is the first result on showing the explicit estimation errors for both ability of users and ground truth for the problem. Also, this paper is the first result on theoretically comparing with the private major voting algorithm.

    
    Di Wang and Jinhui Xu
    Theoretical Computer Science Volume 865, 14 April 2021, Pages 85-98
    


50. [TCS] Differentially Private High Dimensional Sparse Covariance Matrix Estimation [Link] Abstract▼

    In this paper, we study the problem of estimating the covariance matrix under differential privacy, where the underlying covariance matrix is assumed to be sparse and of high dimensions. We propose a new method, called DP-Thresholding, to achieve a non-trivial $\ell_2$-norm based error bound, which is significantly better than the existing ones from adding noise directly to the empirical covariance matrix. We also extend the $\ell_2$-norm based error bound to a general $\ell_w$-norm based one for any $1\leq w\leq \infty$, and show that they share the same upper bound asymptotically. Our approach can be easily extended to local differential privacy. Experiments on the synthetic datasets show consistent results with our theoretical claims.

    
    Di Wang and Jinhui Xu
    Theoretical Computer Science Volume 865, 14 April 2021, Pages 119-130
    



2020

Conference Papers

51. [ICML] On Differentially Private Stochastic Convex Optimization with Heavy-tailed Data [Link] Abstract▼

    In this paper, we consider the problem of designing Differentially Private (DP) algorithms for Stochastic Convex Optimization (SCO) on heavy-tailed data. The irregularity of such data violates some key assumptions used in almost all existing DP-SCO and DP-ERM methods, resulting in failure to provide the DP guarantees. To better understand this type of challenges, we provide in this paper a comprehensive study of DP-SCO under various settings. First, we consider the case where the loss function is strongly convex and smooth. For this case, we propose a method based on the sample-and-aggregate framework, which has an excess population risk of $\tilde{O}(\frac{d^3}{n\epsilon^4})$ (after omitting other factors), where $n$ is the sample size and $d$ is the dimensionality of the data. Then, we show that with some additional assumptions on the loss functions, it is possible to reduce the \textit{expected} excess population risk to $\tilde{O}(\frac{ d^2}{ n\epsilon^2 })$. To lift these additional conditions, we also provide a gradient smoothing and trimming based scheme to achieve excess population risks of $\tilde{O}(\frac{ d^2}{n\epsilon^2})$ and $\tilde{O}(\frac{d^\frac{2}{3}}{(n\epsilon^2)^\frac{1}{3}})$ for strongly convex and general convex loss functions, respectively, \textit{with high probability}. Experiments on both synthetic and real-world datasets suggest that our algorithms can effectively deal with the challenges caused by data irregularity.

    
    Di Wang^*, Hanshen Xiao^*, Srini Devadas and Jinhui Xu (* equal contribution)
    The 37th International Conference on Machine Learning (ICML 2020)


52. [AAAI] Scalable Estimating Stochastic Linear Combination of Non-linear Regressions [Link] Abstract▼

    In this paper we study the problem of estimating stochastic linear combination of non-linear regressions, which has a closed connection with many machine learning and statistical models such as non-linear regressions, Single Index Models, Multi-index Models, Varying Coefficient Index Models and Two-layer Neural Networks. Specifically, we first show that under some mild assumptions, if the variates are multivariate Gaussian, then there is an algorithm whose outputs have $\ell_2$-norm estimation error of $O(\sqrt{\frac{p}{n}})$ with high probability, where $p$ is the dimensionality and $n$ is the size of samples. Moreover, we extend to the bounded sub-Gaussian case by using the zero-bias transformation, which could be seen as a generalization of the classical Stein's lemma. We show that with some additional assumptions there is an algorithm whose outputs have $\ell_\infty$-norm estimation error of $O(\frac{1}{\sqrt{p}}+\sqrt{\frac{p}{n}})$ with high probability. Finally, for both Gaussian and sub-Gaussian cases we propose a scalable algorithm based on sub-sampling method and show that when the sub-sample size is large enough the estimation errors will be almost the same as previous ones. Experimental results for both Gaussian and sub-Gaussian cases support our theoretical results. To our best knowledge, this is the first paper study and provide theoretical guarantees of the stochastic linear combination of non-linear regressions model.

    
    Di Wang^* , Xiangyu Guo^*, Chaowen Guan, Shi Li and Jinhui Xu (* equal contribution)
    The Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI 2020)
    


53. [AAAI] Pairwise Learning with Differential Privacy Guarantees [Link] Abstract▼

    Pairwise learning has received much attention recently as it is more capable of modeling the relative relationship between pairs of samples. Many machine learning tasks can be categorized as pairwise learning, such as AUC maximization and metric learning. Existing techniques for pairwise learning all fail to take into consideration a critical issue in their design, i.e., the protection of sensitive information in the training set. Models learned by such algorithms can implicitly memorize the details of sensitive information, which offers opportunity for malicious parties to infer it from the learned models. To address this challenging issue, in this paper, we propose several differentially private pairwise learning algorithms for both online and offline settings. Specifically, for the online setting, we first introduce a differentially private algorithm (called OnPairStrC) for strongly convex loss functions. Then, we extend this algorithm to general convex loss functions and give another differentially private algorithm (called OnPairC). For the offline setting, we also present two differentially private algorithms (called OffPairStrC and OffPairC) for strongly and general convex loss functions, respectively. These proposed algorithms can not only learn the model effectively from the data but also provide strong privacy protection guarantee for sensitive information in the training set. Extensive experiments on real-world datasets are conducted to evaluate the proposed algorithms and the experimental results support our theoretical analysis.

    
    Mengdi Huai^*, Di Wang^*, Chenglin Miao, Jinhui Xu and Aidong Zhang (* equal contribution)
    Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI 2020)
    


54. [AAAI] Towards Interpretation of Pairwise Learning [Link] Abstract▼

    Recently, there are increasingly more attentions paid to an important family of learning problems called pairwise learning, in which the associated loss functions depend on pairs of instances. Despite the tremendous success of pairwise learning in many real-world applications, the lack of transparency behind the learned pairwise models makes it difficult for users to understand how particular decisions are made by these models, which further impedes users from trusting the predicted results. To tackle this problem, in this paper, we study feature importance scoring as a specific approach to the problem of interpreting the predictions of black-box pairwise models. Specifically, we first propose a novel adaptive Shapley-value-based interpretation method, based on which a vector of importance scores associated with the underlying features of a testing instance pair can be adaptively calculated with the consideration of feature correlations, and these scores can be used to indicate which features make key contributions to the final prediction. Considering that Shapley-value-based methods are usually computationally challenging, we further propose a novel robust approximation interpretation method for pairwise models. This method is not only much more efficient but also robust to data noise. To the best of our knowledge, we are the first to investigate how to enable interpretation in pairwise learning. Theoretical analysis and extensive experiments demonstrate the effectiveness of the proposed methods.

    
    Mengdi Huai, Di Wang, Chenglin Miao and Aidong Zhang
    The Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI 2020)


55. [ECML-PKDD] Escaping Saddle Points of Empirical Risk Privately and Scalably via DP-Trust Region Method [Link] Abstract▼

    It has been shown recently that many non-convex objective/loss functions in machine learning and deep learning are known to be strict saddle. This means that finding a second-order stationary point (i.e., approximate local minimum) and thus escaping saddle points are sufficient for such functions to obtain a classifier with good generalization performance. Existing algorithms for escaping saddle points, however, all fail to take into consideration a critical issue in their designs, that is, the protection of sensitive information in the training set. Models learned by such algorithms can often implicitly memorize the details of sensitive information, and thus offer opportunities for malicious parties to infer it from the learned models. In this paper, we investigate the problem of privately escaping saddle points and finding a second-order stationary point of the empirical risk of non-convex loss function. Previous result on this problem is mainly of theoretical importance and has several issues ({e.g., high sample complexity and non-scalable) which hinder its applicability, especially, in big data. To deal with these issues, we propose in this paper a new method called Differentially Private Trust Region, and show that it outputs a second-order stationary point with high probability and less sample complexity, compared to the existing one. Moreover, we also provide a stochastic version of our method (along with some theoretical guarantees) to make it faster and more scalable. Experiments on benchmark datasets suggest that our methods are indeed more efficient and practical than the previous one.

    
    Di Wang and Jinhui Xu
    The 2020 European Conference on Machine Learning (ECML-PKDD 2020)


56. [BIBM] Global Interpretation for Patient Similarity Learning [Link] Abstract▼

    As an important family of learning problems, pairwise learning has received much attention in recent years. Since pairwise learning involves pairs of instances in its loss function, it is more capable of modeling the relative relationships between instances compared with traditional pointwise learning (e.g., classification). In practice, many machine learning and data mining tasks can be categorized as pairwise learning. Although pairwise learning has achieved tremendous success in many real-world applications, the lack of transparency behind the behavior of the learned pairwise model impedes users from trusting the predicted results, which hampers its further applications in the real world. To tackle this problem, in this paper, we investigate how to enable interpretation in pairwise learning and propose a global interpretation method for pairwise learning. Based on the proposed global interpretation method, we can identify a minimal sufficient subset of data features that are sufficient in themselves to justify the global predictions made by the pairwise model. The identified minimal sufficient feature subset can help us better understand the overall behaviors of the learned pairwise model across different subpopulations of instance pairs. To the best of our knowledge, this is the first work that provides global interpretation for pairwise learning. We also conduct extensive experiments on real-world datasets to evaluate the performance of the proposed global interpretation method.

    
    Mengdi Huai, Chenglin Miao, Jinduo Liu, Di Wang, Jingyuan Chou and Aidong Zhang.
    The 2020 IEEE International Conference on Bioinformatics and Biomedicine (IEEE BIBM 2020)
    Selected as Regular Paper (Acceptance Rate: 19.4%)



Journal Papers

57. [JMLR] Empirical Risk Minimization in the Non-interactive Local Model of Differential Privacy [Link] Abstract▼

    In this paper, we study the Empirical Risk Minimization (ERM) problem in the non-interactive Local Differential Privacy (LDP) model. We first show that if the loss function is $(\infty, T)$-smooth, by using the Bernstein polynomial approximation we can avoid a dependency of the sample complexity, to achieve error $\alpha$, on the exponential of the dimensionality $p$ with base $1/\alpha$ (i.e., $\alpha^{-p}$). This answers a question from (Smith et.al., 2017). Then, we propose player-efficient algorithms with $1$-bit communication complexity and $O(1)$ computation cost for each player. The error bound of these algorithms is asymptotically the same as the original one. With some additional assumptions, we also give an algorithm which is more efficient for the server. Based on different types of polynomial approximations, we propose (efficient) non-interactive locally differential private algorithms for learning the set of k-way marginal queries and the set of smooth queries. Moreover, we study the case of $1$-Lipschitz generalized linear convex loss functions and show that there is an $(\epsilon, \delta)$-LDP algorithm whose sample complexity for achieving error $\alpha$ is only linear in the dimensionality $p$ and quasi-polynomial in other terms. To prove this, we first show that the conclusion holds for the hinge loss function. Then, we extend the result to any $1$-Lipschitz generalized linear convex loss functions by showing that every such a function can be approximated by a linear combination of hinge loss functions and some linear functions. Our results use a polynomial of inner product approximation technique. Then we apply our technique to the Euclidean median problem and show that its sample complexity needs only to be quasi-polynomial in $p$, which is the first result with a sub-exponential sample complexity in $p$ for non-generalized linear loss functions.

    
    Di Wang, Marco Gaboardi, Adam Smith and Jinhui Xu
    Journal of Machine Learning Research, Volume 21, 200 (2020), Pages 1-39


58. [MLJ] Robust High Dimensional Expectation Maximization Algorithm via Trimmed Hard Thresholding [Link] Abstract▼

    In this paper, we study the problem of estimating latent variable models with arbitrarily corrupted samples in the high dimensional case, i.e., $d\gg n$, where the underlying parameter is sparse. Specifically, we propose a method called Trimmed (Gradient) Expectation Maximization which attaches a trimming gradients and hard thresholding step to the Expectation step (E-step) and Maximization step (M-step), respectively. Particularly, under some mild assumptions, with an appropriate initialization, we show that the algorithm is corruption-free and converges to the (near) optimal statistical rate geometrically when the fraction of corruption samples satisfies $\alpha\leq O(\frac{1}{\sqrt{n}})$. Moreover, we implement our general framework to three canonical examples: mixture of Gaussians, mixture of regressions and linear regression with missing covariates. Experiments also support our theoretical analysis.

    
    Di Wang^*, Xiangyu Guo^*, Shi Li and Jinhui Xu (* equal contribution)
    Machine Learning, 109, 2283-2311 (2020)
    


59. [TCS] Tight Lower Bound of Locally Differentially Private Sparse Covariance Matrix Estimation [Link] Abstract▼

    In this paper, we study the sparse covariance matrix estimation problem in the local differential privacy model, and give a lower bound of $\Omega(\frac{s^2\log p}{n\epsilon^2})$ on the $\epsilon$ non-interactive private minimax risk in the metric of squared spectral norm, where $s$ is the row sparsity of the underlying covariance matrix, $n$ is the sample size, and $p$ is the dimensionality of the data. We show that the lower bound is actually tight, as it matches a previous upper bound.Our main technique for achieving this lower bound is a general framework, called General Private Assouad Lemma, which is a considerable generalization of the previous private Assouad lemma and can be used as a general method for bounding the private minimax risk of matrix-related estimation problems.

    
    Di Wang and Jinhui Xu
    Theoretical Computer Science, Volume 815, 2 May 2020, Pages 47-59
    


60. [TCS] Principal Component Analysis in the Local Differential Privacy Model [Link] Abstract▼

    In this paper, we study the Principal Component Analysis (PCA) problem under the (distributed) non-interactive local differential privacy model. For the low dimensional case (i.e., $p \ll n$), we show the optimal rate of $\Theta(\frac{kp}{n\epsilon^2})$ (omitting the eigenvalue terms) for the private minimax risk of the $k$-dimensional PCA using the squared subspace distance as the measurement, where $n$ is the sample size and $\epsilon$ is the privacy parameter. For the high dimensional (i.e., $p\gg n$) row sparse case, we first give a lower bound of $\Omega(\frac{ks\log p }{n\epsilon^2})$ on the private minimax risk, where $s$ is the underlying sparsity parameter. Then we provide an efficient algorithm to achieve the upper bound of $O(\frac{s^2\log p}{n\epsilon^2})$. Experiments on both synthetic and real world datasets confirm our theoretical guarantees.

    
    Di Wang and Jinhui Xu
    Theoretical Computer Science, Volume 809, 24 February 2020, Pages 296-312
    


61. [Neurocomputing] Estimating Stochastic Linear Combination of Non-linear Regressions Efficiently and Scalably [Link] Abstract▼

    Recently, many machine learning and statistical models such as non-linear regressions, the Single Index, Multi-index, Varying Coefficient Index Models and Two-layer Neural Networks can be reduced to or be seen as a special case of a new model which is called the \textit{Stochastic Linear Combination of Non-linear Regressions} model. However, due to the high non-convexity of the problem, there is no previous work study how to estimate the model. In this paper, we provide the first study on how to estimate the model efficiently and scalably. Specifically, we first show that with some mild assumptions, if the variate vector $x$ is multivariate Gaussian, then there is an algorithm whose output vectors have $\ell_2$-norm estimation errors of $O(\sqrt{\frac{p}{n}})$ with high probability, where $p$ is the dimension of $x$ and $n$ is the number of samples. The key idea of the proof is based on an observation motived by the Stein's lemma. Then we extend our result to the case where $x$ is bounded and sub-Gaussian using the zero-bias transformation, which could be seen as a generalization of the classic Stein's lemma. We also show that with some additional assumptions there is an algorithm whose output vectors have $\ell_\infty$-norm estimation errors of $O(\frac{1}{\sqrt{p}}+\sqrt{\frac{p}{n}})$ with high probability. We also provide a concrete example to show that there exists some link function which satisfies the previous assumptions. Finally, for both Gaussian and sub-Gaussian cases we propose a faster sub-sampling based algorithm and show that when the sub-sample sizes are large enough then the estimation errors will not be sacrificed by too much. Experiments for both cases support our theoretical results. To the best of our knowledge, this is the first work that studies and provides theoretical guarantees for the stochastic linear combination of non-linear regressions model.

    
    Di Wang^* , Xiangyu Guo^* , Chaowen Guan, Shi Li and Jinhui Xu (* equal contribution)
    Neurocomputing, Volume 399, 25 July 2020, Pages 129-140



2019

Conference Papers

62. [ICML] Differentially Private Empirical Risk Minimization with Non-convex Loss Functions [Link] Abstract▼

    We study the problem of Empirical Risk Minimization (ERM) with (smooth) non-convex loss functions under the differential-privacy (DP) model. Existing approaches for this problem mainly adopt gradient norms to measure the error, which in general cannot guarantee the quality of the solution. To address this issue, we first study the expected excess empirical (or population) risk, which was primarily used as the utility to measure the quality for convex loss functions. Specifically, we show that the excess empirical (or population) risk can be upper bounded by $\tilde{O}(\frac{d\log (1/\delta)}{\log n\epsilon^2})$ in the $(\epsilon, \delta)$-DP settings, where $n$ is the data size and $d$ is the dimensionality of the space. The $\frac{1}{\log n}$ term in the empirical risk bound can be further improved to $\frac{1}{n^{\Omega(1)}}$ (when $d$ is a constant) by a highly non-trivial analysis on the time-average error. To obtain more efficient solutions, we also consider the connection between achieving differential privacy and finding approximate local minimum. Particularly, we show that when the size $n$ is large enough, there are $(\epsilon, \delta)$-DP algorithms which can find an approximate local minimum of the empirical risk with high probability in both the constrained and non-constrained settings. These results indicate that one can escape saddle points privately.

    
    Di Wang, Changyou Chen and Jinhui Xu
    The 36th International Conference on Machine Learning (ICML 2019)


63. [ICML] On Sparse Linear Regression in the Local Differential Privacy Model [Link] Abstract▼

    In this paper, we study the sparse linear regression problem under the Local Differential Privacy (LDP) model. We first show that polynomial dependency on the dimensionality $p$ of the space is unavoidable for the estimation error in both non-interactive and sequential interactive local models, if the privacy of the whole dataset needs to be preserved. Similar limitations also exist for other types of error measurements and in the relaxed local models. This indicates that differential privacy in high dimensional space is unlikely achievable for the problem. With the understanding of this limitation, we then present two algorithmic results. The first one is a sequential interactive LDP algorithm for the low dimensional sparse case, called Locally Differentially Private Iterative Hard Thresholding (LDP-IHT), which achieves a near optimal upper bound. This algorithm is actually rather general and can be used to solve quite a few other problems, such as (Local) DP-ERM with sparsity constraints and sparse regression with non-linear measurements. The second one is for the restricted (high dimensional) case where only the privacy of the responses (labels) needs to be preserved. For this case, we show that the optimal rate of the error estimation can be made logarithmically depending on $p$ (i.e., $\log p$) in the local model, where an upper bound is obtained by a label-privacy version of LDP-IHT. Experiments on real world and synthetic datasets confirm our theoretical analysis.

    
    Di Wang and Jinhui Xu
    The 36th International Conference on Machine Learning (ICML 2019)
    Selected as Long Talk (Acceptance Rate: 140/3424= 4.1%)
    NeurIPS 2018 Workshop on Privacy Preserving Machine Learning
    


64. [NeurIPS] Facility Location Problem in Differential Privacy Model Revisited [Link] Abstract▼

    In this paper we study the uncapacitated facility location problem in the model of differential privacy (DP) with uniform facility cost. Specifically, we first show that, under the \emph{hierarchically well-separated tree (HST) metrics} and the super-set output setting that was introduced in (Gupta et al 2010), there is an $\epsilon$-DP algorithm that achieves an $O(\frac{1}{\epsilon})$ (expected multiplicative) approximation ratio; this implies an $O(\frac{\log n}{\epsilon})$ approximation ratio for the general metric case, where $n$ is the size of the input metric. These bounds improve the best-known results given by (Gupta et al 2010). In particular, our approximation ratio for HST-metrics is independent of $n$, and the ratio for general metrics is independent of the aspect ratio of the input metric. On the negative side, we show that the approximation ratio of any $\epsilon$-DP algorithm is lower bounded by $\Omega(\frac{1}{\sqrt{\epsilon}})$, even for instances on HST metrics with uniform facility cost, under the super-set output setting. The lower bound shows that the dependence of the approximation ratio for HST metrics on $\epsilon$ can not be removed or greatly improved. Our novel methods and techniques for both the upper and lower bound may find additional applications.

    
    [alphabetic order] Yunus Esencayi, Marco Gaboardi, Shi Li and Di Wang
    Conference on Neural Information Processing Systems (NIPS/NeurIPS), 2019


65. [ALT] Noninteractive Locally Private Learning of Linear Models via Polynomial Approximations [Link] Abstract▼

    In this paper, we study the Empirical Risk Minimization problem in the non-interactive Local Differential Privacy (LDP) model. First, we show that for the hinge loss function, there is an $(\epsilon, \delta)$-LDP algorithm whose sample complexity for achieving an error of $\alpha$ is only linear in the dimensionality $p$ and quasi-polynomial in other terms. Then, we extend the result to any $1$-Lipschitz generalized linear convex loss functions by showing that every such function can be approximated by a linear combination of hinge loss functions and some linear functions. Finally, we apply our technique to the Euclidean median problem and show that its sample complexity needs only to be quasi-polynomial in $p$, which is the first result with a sub-exponential sample complexity in $p$ for non-generalized linear loss functions. Our results are based on a technique, called polynomial of inner product approximation, which may be applicable to other problems.

    
    Di Wang, Adam Smith and Jinhui Xu
    The 30th International Conference on Algorithmic Learning Theory (ALT 2019)


66. [AAAI] Differentially Private Empirical Risk Minimization with Smooth Non-convex Loss Functions: A Non-stationary View [Link] Abstract▼

    In this paper, we study the Differentially Private Empirical Risk Minimization (DP-ERM) problem with non-convex loss functions and give several upper bounds for the utility in different settings. We first consider the problem in low-dimensional space. For DP-ERM with non-smooth regularizer, we generalize an existing work by measuring the utility using $\ell_2$ norm of the projected gradient. Also, we extend the error bound measurement, for the first time, from empirical risk to population risk by using the expected $\ell_2$ norm of the gradient. We then investigate the problem in high dimensional space, and show that by measuring the utility with Frank-Wolfe gap, it is possible to bound the utility by the Gaussian Width of the constraint set, instead of the dimensionality $p$ of the underlying space. We further demonstrate that the advantages of this result can be achieved by the measure of $\ell_2$ norm of the projected gradient. A somewhat surprising discovery is that although the two kinds of measurements are quite different, their induced utility upper bounds are asymptotically the same under some assumptions. We also show that the utility of some special non-convex loss functions can be reduced to a level (i.e., depending only on $\log p$) similar to that of convex loss functions. Finally, we test our proposed algorithms on both synthetic and real world datasets and the experimental results confirm our theoretical analysis.

    
    Di Wang and Jinhui Xu
    The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI 2019)
    Selected as Oral Presentation (Acceptance Rate: 460/7095=6.5%)


67. [IJCAI] Lower Bound of Locally Differentially Private Sparse Covariance Matrix Estimation [Link] Abstract▼

    In this paper, we study the sparse covariance matrix estimation problem in the local differential privacy model, and give a non-trivial lower bound on the non-interactive private minimax risk in the metric of squared spectral norm. We show that the lower bound is actually tight, as it matches a previous upper bound. Our main technique for achieving this lower bound is a general framework, called General Private Assouad Lemma, which is a considerable generalization of the previous private Assouad lemma and can be used as a general method for bounding the private minimax risk of matrix-related estimation problems.

    
    Di Wang and Jinhui Xu
    The 28th International Joint Conference on Artificial Intelligence (IJCAI 2019)


68. [IJCAI] Principal Component Analysis in the Local Differential Privacy Model [Link] Abstract▼

    In this paper, we study the Principal Component Analysis (PCA) problem under the (distributed) non-interactive local differential privacy model. For the low dimensional case, we show the optimal ratefor the private minimax risk of the kdimensional PCA using the squared subspace distance as the measurement. For the high dimensional row sparse case, we first give a lower bound on the private minimax risk, . Then we provide an efficient algorithm to achieve a near optimal upper bound. Experiments on both synthetic and real world datasets confirm the theoretical guarantees of our algorithms.

    
    Di Wang and Jinhui Xu
    The 28th International Joint Conference on Artificial Intelligence (IJCAI 2019)


69. [IJCAI] Privacy-aware Synthesizing for Crowdsourced Data [Link] Abstract▼

    The prevalence of the World Wide Web enables the data collectors to easily share their data that are collected from a large crowd of users with the public. Although releasing crowdsourced data on the web brings many benefits to the data analyzers to conduct statistical analysis, it may violate crowd users' data privacy, which has been a major obstacle to release the crowdsourced data. A potential way to address this problem is to employ the traditional differential privacy-based mechanisms and perturb the data with some noise before publishing them. However, such noise perturbation mechanisms offer little data utility because crowdsourced datasets contain massive amounts of conflicting data records with large domains. In particular, this degraded utility results from two respects: firstly, the originally collected crowdsourced data usually contain conflicting data; secondly, the noise needed to guarantee differential privacy may be proportional to the number of data records or the domain of the input data, which renders the released crowdsourced data useless. To address the above challenges, we propose a novel privacy-aware synthesizing method for crowdsourced data. In this method, the data collectors first learn the underlying distributions of the crowdsourced data through taking each user's fine grained reliability degrees into account. Then, a set of candidate synthetics are sampled from the learned distributions. Finally, these sampled candidate synthetics are subjected to privacy tests, and only those candidate synthetics which pass privacy tests are allowed to be safely released. The proposed method not only provides strong privacy protection for individual users but also generates high utility synthetic data. The desirable performance of the proposed method is verified via theoretical analysis and extensive experiments conducted on both real-world and synthetic datasets.

    
    Mengdi Huai, Di Wang, Chenglin Miao, Jinhui Xu, Aidong Zhang
    The 28th International Joint Conference on Artificial Intelligence (IJCAI 2019)


70. [CISS] Estimating Sparse Covariance Matrix Under Differential Privacy via Thresholding [Link] Abstract▼

    In this paper, we study the problem of estimating the covariance matrix under differential privacy, where the underlying covariance matrix is assumed to be sparse and of high dimensions. We propose a new method, called DP-Thresholding, to achieve a non-trivial $\ell_2$-norm based error bound, which is significantly better than the existing ones from adding noise directly to the empirical covariance matrix. Experiments on the synthetic datasets show consistent results with our theoretical claims.

    
    Di Wang, Jinhui Xu and Yang He
    The 53rd Annual Conference on Information Sciences and Systems (CISS 2019)



Journal Papers

71. [Neurocomputing] Faster Large Scale Constrained Linear Regression via Two-Step Preconditioning [Link] Abstract▼

    In this paper, we study the large scale constrained linear regression problem and propose a two-step preconditioning method, which is based on some recent developments on random projection, sketching techniques and convex optimization methods. Combining the method with (accelerated) mini-batch SGD, we can achieve an approximate solution with a time complexity lower than that of the state-of-the-art techniques for the low precision case. Our idea can also be extended to the high precision case, which gives an alternative implementation to the Iterative Hessian Sketch (IHS) method with significantly improved time complexity. Experiments on benchmark and synthetic datasets suggest that our methods indeed outperform existing ones considerably in both the low and high precision cases.

    
    Di Wang and Jinhui Xu
    Neurocomputing, Volume 364, 28 October 2019, Pages 280-296



2018

Conference Papers

72. [NeurIPS] Empirical Risk Minimization in Non-interactive Local Differential Privacy Revisited [Link] Abstract▼

    In this paper, we revisit the Empirical Risk Minimization problem in the non-interactive local model of differential privacy. In the case of constant or low dimensions ($p\ll n$), we first show that if the loss function is $(\infty, T)$-smooth, we can avoid a dependence of the sample complexity, to achieve error $\alpha$, on the exponential of the dimensionality $p$ with base $1/\alpha$ (i.e., $\alpha^{-p}$), which answers a question in (Smith et al 2017). Our approach is based on polynomial approximation. Then, we propose player-efficient algorithms with $1$-bit communication complexity and $O(1)$ computation cost for each player. The error bound is asymptotically the same as the original one. With some additional assumptions, we also give an efficient algorithm for the server. In the case of high dimensions ($n\ll p$), we show that if the loss function is a convex generalized linear function, the error can be bounded by using the Gaussian width of the constrained set, instead of $p$, which improves the one in Smith et al. Our techniques can be extended to some related problems, such as $k$-way marginal queries and smooth queries.

    
    Di Wang, Marco Gaboardi and Jinhui Xu
    Conference on Neural Information Processing Systems (NIPS/NeurIPS), 2018


73. [AAAI ] Large Scale Constrained Linear Regression Revisited: Faster Algorithms via Preconditioning [Link] Abstract▼

    In this paper, we revisit the large-scale constrained linear regression problem and propose faster methods based on some recent developments in sketching and optimization. Our algorithm combines mini-batch SGD with a new method called two-step preconditioning to achieve an $\epsilon$-accuracy solution for the low precision case, and has a lower time complexity than the state-of-the-art techniques. Our idea can also be extended to the high precision case, which gives an alternative implementation to the Iterative Hessian Sketch (IHS) method with significantly improved time complexity. Experiments on benchmark and synthetic datasets suggest that our methods indeed outperform existing ones considerably in both the low and high precision cases.

    
    Di Wang and Jinhui Xu
    The Thirty-Second AAAI Conference on Artificial Intelligence (AAAI 2018)
    Selected as Oral Presentation (Acceptance Rate: 411/3800=10.8%)


74. [GlobalSip] Differentially Private Sparse Inverse Covariance Estimation [Link] Abstract▼

    In this paper, we give the first study of sparse inverse covariance estimation problem under differential privacy. Firstly, we propose an $\epsilon$-differentially private algorithm via output perturbation, which is based on the sensitivity of the optimization problem and Wishart mechanism. Based on the idea of that, we propose a general covariance perturbation method, and then for $\epsilon$-differential privacy, we analyze Laplacian and Wishart mechanisms, for $(\epsilon,\delta)$-differential privacy we analyze Gaussian and Wishart mechanisms. Moreover, we extend the covariance perturbation algorithm to distributed setting and local differential privacy. Experiments on synthetic and benchmark datasets are also support our theoretical analysis.

    
    Di Wang, Mengdi Huai and Jinhui Xu
    2018 6th IEEE Global Conference on Signal and Information Processing (2018 GlobalSip)
    Selected as Oral Presentation



2017

Conference Papers

75. [NeurIPS] Differentially Private Empirical Risk Minimization Revisited: Faster and More General [Link] Abstract▼

    In this paper we study the differentially private Empirical Risk Minimization (ERM) problem in different settings. For smooth (strongly) convex loss function with or without (non)-smooth regularization, we give algorithms that achieve either optimal or near optimal utility bounds with less gradient complexity compared with previous work. For ERM with smooth convex loss function in high-dimensional ($p\gg n$) setting, we give an algorithm which achieves the upper bound with less gradient complexity than previous ones. At last, we generalize the expected excess empirical risk from convex loss functions to non-convex ones satisfying the Polyak-Lojasiewicz condition and give a tighter upper bound on the utility than the one in (Zhang et al 2017).

    
    Di Wang, Minwei Ye and Jinhui Xu
    Conference on Neural Information Processing Systems (NIPS/NeurIPS), 2017