基于分组蒸馏的可验证隐私保护公平个性化联邦学习方法

A Verifiable Privacy-Preserving Fair Personalized Federated Learning Method Based on Group Distillation

摘要: 个性化联邦学习(Personalized Federated Learning, PFL)允许每个用户在共享全局模型的基础上训练个性化模型，显著提升了模型在异构数据环境下的性能。然而也面临数据隐私泄露、客户端之间模型性能失衡以及聚合结果可信性等方面的挑战。为了解决上述问题，文章设计了基于用户数据相似性的层次聚类分组机制与跨组知识蒸馏机制，在不泄露原始梯度的前提下实现相似数据用户的精准聚类，将数据特征相近的用户归为同一组，共同开展联邦训练，实现数据层面的个性化；并缩小用户之间的模型性能差距，保障了系统的公平性。在隐私保护方面，设计了基于掩码的安全聚合机制，确保各用户上传的梯度参数在传输与聚合过程中始终处于隐私保护状态。同时，设计了验证机制，使诚实用户在接收聚合结果后能够独立校验其正确性。实验结果表明，本方案满足了高级别隐私保护要求，确保了用户之间的公平性，并且模型准确率优于现有方案。

Abstract: Personalized Federated Learning (PFL) allows each user to train a personalized model based on a shared global model, significantly improving model performance in heterogeneous data environments. However, it also faces challenges such as data privacy leakage, imbalance in model performance among clients, and the trustworthiness of aggregation results. To address these issues, a hierarchical clustering grouping mechanism based on user data similarity and a cross-group knowledge distillation mechanism were designed. This enables precise clustering of users with similar data without revealing original gradients, allowing users with similar data to train together in the same group for federated training, achieving data-level personalization; it also reduces the performance gap between users, ensuring system fairness. In terms of privacy protection, a mask-based secure aggregation mechanism was designed to ensure that the gradient parameters uploaded by users remain protected throughout transmission and aggregation. At the same time, a verification mechanism was designed so that honest users can independently verify the correctness of the aggregated results once received. Experimental results show that this solution meets high-level privacy protection requirements, ensures fairness among users, and achieves model accuracy superior to existing solutions.