计算机技术与发展2025，Vol.35Issue(11)：102-113,12.DOI:10.20165/j.cnki.ISSN1673-629X.2025.0174
基于强化学习的自适应差分隐私在线联邦学习算法
Reinforcement Learning-based Adaptive Online Federated Learning Algorithm with Differential Privacy
摘要
Abstract
In the current era of rapid growth in information technology,online federated learning based on real-time user privacy data can establish effective machine learning models,which is beneficial to people's lives.However,current federated learning still faces the problem of low communication efficiency and low usability when dealing with user privacy data streams.In response to the above issues,based on the w-event level differential privacy model,we propose an adaptive differential privacy online federated learning algorithm based on reinforcement learning,which achieves dynamic control of online federated learning by integrating reinforcement learning.In this process,we adopt the deep Q-network algorithm in deep reinforcement learning and combines it with dimensionality reduction algorithm to transform the federated learning model into an embedded representation.Input embedded representations and environmental information together into the Q-network to dynamically regulate the user sampling rate and aggregation interval of the central server in the online federated learning process.The dynamic privacy allocation algorithm of the client is combined with the dynamic sampling algorithm of the central server to ensure that the locally transmitted data from the client satisfiesw-event level differential privacy.The experimental comparison results on public datasets show that the proposed algorithm exhibits significant advantages in accuracy and com-munication efficiency compared to traditional algorithms.关键词
差分隐私/联邦学习/在线学习/隐私保护/强化学习Key words
differential privacy/federated learning/online learning/privacy protection/reinforcement learning分类
计算机与自动化引用本文复制引用
陆光前,曾晓雯,张泽泉,胡健,何映军,李好..基于强化学习的自适应差分隐私在线联邦学习算法[J].计算机技术与发展,2025,35(11):102-113,12.基金项目
云南电网科技项目(YNKJXM20222531) （YNKJXM20222531）
国家自然科学基金(62172329,61802298) （62172329,61802298）
