基于差分隐私的分段裁剪联邦学习算法

Segmental tailoring federated learning algorithm based on differential privacy

为解决现有的差分隐私联邦学习算法中使用固定的裁剪阈值和噪声尺度进行训练,从而导致数据隐私泄露、模型精度较低的问题,提出了一种基于差分隐私的分段裁剪联邦学习算法。首先,根据客户端的隐私需求分为隐私需求高和低。对于高隐私需求用户使用自适应裁剪来动态裁剪梯度,而低隐私需求用户则采用比例裁剪。其次根据裁剪后阈值大小自适应地添加噪声尺度。通过实验分析可得,该算法可以更好地保护隐私数据,同时通信代价也低于ADP-FL和DP-FL算法,并且与ADP-FL和DP-FL相比,模型准确率分别提高了2.25%和4.41%。

To solve the problems caused by using fixed cropping thresholds and noise scales for training in existing differential privacy federated learning algorithms,such as data privacy leakage and low model accuracy,the paper proposed a segmented cropping federated learning algorithm based on differential privacy.Firstly,the clients divided the privacy requirements into high and low privacy demands.For users with high privacy demands,it employed adaptive clipping to dynamically clip the gradients.Conversely,for users with low privacy demands,it adopted proportional clipping.Secondly,the clients adaptively added noise scales based on the size of the clipped threshold.The experimental analysis shows that this algorithm effectively safeguards privacy data,while reducing communication costs compared to ADP-FL and DP-FL algorithms.Additionally,it achieves an improvement in model accuracy by 2.25%and 4.41%compared to ADP-FL and DP-FL respectively.