基于自适应噪声和动态加权的联邦学习算法  

Federated learning scheme based on adaptive noise and dynamic weighting

作　　者：王红林[1] 薛珊 朱丞 Wang Honglin;Xue Shan;Zhu Cheng(School of Artificial Intelligence(College of Future Technology),Nanjing University of Information Science&Technology,Nanjing 210044,China;School of Computer science,Nanjing University of Information Science&Technology,Nanjing 210044,China;College of Electrical&Computer Engineering,University of Illinois at Urbana Champaign,Urbana IL 61801,USA)

机构地区：[1]南京信息工程大学人工智能学院(未来技术学院),南京210044 [2]南京信息工程大学计算机学院,南京210044 [3]伊利诺伊大学厄巴纳-香槟分校电子与计算机工程学院,美国厄巴纳61801

出　　处：《计算机应用研究》2025年第3期749-754,共6页Application Research of Computers

基　　金：国家自然科学基金委员会青年项目(62101275,62101274)。

摘　　要：将差分隐私应用于联邦学习是保护训练数据隐私的有效方法之一,但在现有的算法中,添加固定噪声进行模型训练会导致模型精度不高、数据隐私泄露的问题。为此,提出了一种基于自适应噪声和动态加权的联邦学习算法(DP-FedANAW)。首先,考虑到梯度的异质性,该算法为每个客户端预测当前轮次梯度范数,获得裁剪阈值,为其进行不同轮次自适应裁剪梯度,从而实现自适应调整噪声;其次,为了进一步提高模型的训练效率,该算法还提出了一种将客户端贡献度与数据量相结合的动态加权模型聚合方法。实验结果表明,该算法在满足差分隐私的前提下,与DP-FL和其他两个自适应噪声的算法相比,不仅准确率提高了5.03、2.94和2.85百分点,而且训练轮次整体提高了5~40轮。Applying differential privacy to federated learning is one of the effective methods to protect the privacy of training data,but adding fixed noise to model training in existing algorithms will lead to the problem of low model accuracy and data privacy leakage.Therefore,this paper proposed a federation learning algorithm based on adaptive noise and dynamic weighting.Firstly,considering the heterogeneity of the gradient,the algorithm predicted the current round gradient norm for each client,obtains the clipping threshold,and performed adaptive clipping gradients for different rounds to achieve adaptive noise adjustment.Secondly,in order to further improve the efficiency of model training,the algorithm also proposed a dynamic weighted model aggregation method combining client contribution and data volume.The experimental results show that this algorithm not only improves the accuracy by 5.03,2.94 and 2.85 percent points,but also improves the training rounds by about 5~40 rounds compared with DP-FL and the other two adaptive noise algorithms under the premise of differential privacy.

关 键 词：联邦学习 差分隐私 自适应噪声 自适应裁剪 动态加权 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]