| 摘要: |
| 联邦学习允许众多客户端利用其本地数据联合训练模型而不暴露各方的真实数据, 与传统机器学习方法相比, 避免了数据迁移导致的数据泄露和滥用问题. 然而, 在实际应用中, 客户端可能具有异构的数据分布和系统功能, 这会导致模型性能和训练效率下降. 通过选择一个“良好的”客户端子集作为联邦学习参与方可以有效提高全局模型的性能和收敛速度. 而一些研究者发现, 恶意敌手可以利用客户端的本地训练损失或梯度等一些相关信息来推断其隐私数据, 目前的异构联邦学习参与方选择方案并没有应对这种隐私泄露风险的解决方法. 为此, 设计了一种基于安全多方计算的隐私保护异构联邦学习参与方选择协议, 利用3PC秘密共享技术来保证训练过程的数据隐私以及联合模型的准确度, 同时还提出了一个安全top-k搜索协议来避免参与方选择过程中泄露任何隐私信息. 对协议的安全性进行分析, 证明了该协议可以满足安全需求, 并且开展相关实验, 实验结果表明相比于未使用隐私保护的异构联邦学习方案, 所提出方案的各方的平均计算与通信时间开销仅增加了2.09%. |
| 关键词: 隐私保护 联邦学习 参与方选择 数据异构性 安全多方计算 |
| DOI：10.13328/j.cnki.jos.007622 |
| 分类号:TP18 |
| 基金项目:国家科技创新2030—重大项目(2021ZD0112802); 国家自然科学基金(62332007, 62572214, 62072215, U23A20303, U22B2028, 62302194, 62573201); 西藏自治区科技重大专项(XZ202201ZD0006G); 深圳市科技重大专项(KJZD20240903101108011); 广州市基础与应用基础项目(2024A03J0405, 2024A04J3458); 广东省数据安全与隐私保护重点实验室(2023B1212060036) |
|
| Privacy-preserving Participant Selection Based on Secure Multi-party Computation in Heterogeneous Federated Learning |
|
LIU Teng-Fei1, YANG An-Jia1, WENG Jian1, CHEN Min-Rong2, LIU Yi1, ZENG Huang3
|
|
1.College of Cyber Security, Jinan University, Guangzhou 511443, China;2.School of Computer Science, South China Normal University, Guangzhou 510631, China;3.Air Force Early Warning College, Wuhan 430019, China
|
| Abstract: |
| Federated learning enables numerous clients to collaboratively train a model using their local data without exposing the raw data of individual parties, thereby avoiding data leakage and misuse caused by data migration in traditional machine learning paradigms. However, in practical scenarios, clients often exhibit heterogeneous data distributions and diverse system capabilities, which degrade model performance and training efficiency. Selecting a high-quality subset of clients as participants in federated learning can effectively improve global model performance and accelerate convergence. Nevertheless, existing studies show that malicious adversaries can exploit information such as local training losses or gradients to infer sensitive private data, while current participant selection strategies for heterogeneous federated learning fail to adequately address such privacy leakage risks. To tackle this challenge, this study designs a privacy-preserving participant selection protocol for heterogeneous federated learning based on secure multi-party computation. By leveraging three-party computation (3PC) secret-sharing techniques, the proposed protocol ensures data privacy during training while maintaining the accuracy of the jointly trained model. Furthermore, a secure top-k search protocol is introduced to prevent privacy leakage during the participant selection phase. The security of the proposed protocol is formally analyzed, demonstrating that the required security properties are satisfied. Experimental results indicate that, compared with heterogeneous federated learning schemes without privacy preservation, the proposed approach increases the average computational and communication overhead across all parties by 2.09%. |
| Key words: privacy-preserving federated learning (FL) participant selection data heterogeneity secure multi-party computation |
