A5014906752- Privacy-Preserving Technologies in Data
- Recommender Systems and Techniques
- Stochastic Gradient Optimization Techniques
- Auction Theory and Applications
- Privacy, Security, and Data Protection
- Cryptography and Data Security
- Consumer Market Behavior and Pricing
- Blockchain Technology Applications and Security
- Probability and Risk Models
- Data Quality and Management
- Advanced Authentication Protocols Security
- Advanced Malware Detection Techniques
- IPv6, Mobility, Handover, Networks, Security
- Peer-to-Peer Network Technologies
- Economic and Environmental Valuation
- Computational Drug Discovery Methods
- Ethics and Social Impacts of AI
- Cardiac electrophysiology and arrhythmias
- Metabolomics and Mass Spectrometry Studies
- Advanced Bandit Algorithms Research
- Machine Learning and Data Classification
- Imbalanced Data Classification Techniques
- Psychological and Educational Research Studies
- Distributed Sensor Networks and Detection Algorithms
- Advanced Graph Neural Networks
Zhejiang Lab
2023-2024
Zhejiang University of Science and Technology
2024
Zhejiang University
2024
Tsinghua University
2021-2023
Federated learning is a collaborative machine framework where multiple clients jointly train global model. To mitigate communication overhead, it common to select subset of for participation in each training round. However, existing client selection strategies often rely on fixed number throughout all rounds, which may not be the optimal choice balancing efficiency and model performance. Moreover, these approaches typically evaluate solely based their performances one single round,...
We study convex optimization problems under differential privacy (DP). With heavy-tailed gradients, existing works achieve suboptimal rates. The main obstacle is that gradient estimators have tail property, resulting in a superfluous factor of d the union bound. In this paper, we explore algorithms achieving optimal rates DP with gradients. Our first method simple clipping approach. Under bounded p-th order moments n samples, it achieves minimax population risk epsilon less than 1/d. then...
Deep reinforcement learning (DRL) is widely applied to safety-critical decision-making scenarios. However, DRL vulnerable backdoor attacks, especially action-level backdoors, which pose significant threats through precise manipulation and flexible activation, risking outcomes like vehicle collisions or drone crashes. The key distinction of backdoors lies in the utilization reward function associate triggers with target actions. Nevertheless, existing studies typically rely on functions fixed...
With the continuous development of large language models (LLMs), transformer-based have made groundbreaking advances in numerous natural processing (NLP) tasks, leading to emergence a series agents that use LLMs as their control hub. While achieved success various they face security and privacy threats, which become even more severe agent scenarios. To enhance reliability LLM-based applications, range research has emerged assess mitigate these risks from different perspectives. help...
We face choice problems every day, where we have to choose from options with multiple conflicting attributes, and learning a user's personal preference in is particularly important. For example, E-commerce applications, this study may help make recommendations sales prediction, offer important guidelines on sellers' pricing discount strategies. Prior works either do not consider the competition among different options, or cannot handle convex hull problem user selects an option that regarded...
Inter-temporal choices involve making decisions that require weighing costs in the present against benefits future. One specific type of inter-temporal choice is decision between purchasing an individual item at full price or opting for a bundle including discounted price. Previous works assume users have accurate expectations factors involved these decisions. However, reality, users’ perceptions are often biased, leading to irrational and suboptimal decision-making. In this work, we focus...
Privacy protection of users' entire contribution samples is important in distributed systems. The most effective approach the two-stage scheme, which finds a small interval first and then gets refined estimate by clipping into interval. However, operation induces bias, serious if sample distribution heavy-tailed. Besides, users with large local sizes can make sensitivity much larger, thus method not suitable for imbalanced users. Motivated these challenges, we propose Huber loss minimization...
Label differential privacy (DP) is a framework that protects the of labels in training datasets, while feature vectors are public. Existing approaches protect by flipping them randomly, and then train model to make output approximate privatized label. However, as number classes $K$ increases, stronger randomization needed, thus performances these methods become significantly worse. In this paper, we propose vector approximation approach, which easy implement introduces little additional...
User-level privacy is important in distributed systems. Previous research primarily focuses on the central model, while local models have received much less attention. Under user-level DP strictly stronger than item-level one. However, under relationship between and LDP becomes more complex, thus analysis crucially different. In this paper, we first analyze mean estimation problem then apply it to stochastic optimization, classification, regression. particular, propose adaptive strategies...
Backdoor attacks have attracted wide attention from academia and industry due to their great security threat deep neural networks (DNN). Most of the existing methods propose conduct backdoor by poisoning training dataset with different strategies, so it's critical identify poisoned samples then train a clean model on unreliable in context defending attacks. Although numerous countermeasure researches are proposed, inherent weaknesses render them limited practical scenarios, such as...
Client selection significantly affects the system convergence efficiency and is a crucial problem in federated learning. Existing methods often select clients by evaluating each round individually overlook necessity for long-term optimization, resulting suboptimal performance potential fairness issues. In this study, we propose novel client strategy designed to emulate achieved with full participation. single round, minimizing gradient-space estimation error between subset set. multi-round...
We study convex optimization problems under differential privacy (DP). With heavy-tailed gradients, existing works achieve suboptimal rates. The main obstacle is that gradient estimators have tail properties, resulting in a superfluous factor of $d$ the union bound. In this paper, we explore algorithms achieving optimal rates DP with gradients. Our first method simple clipping approach. Under bounded $p$-th order moments $n$ samples, it achieves...