基于个性化本地动量的自适应差分隐私联邦学习OA

Federated learning(FL)is a distributed machine learning paradigm that enables multiple de-vices or organizations to collaboratively train models by sharing training parameters instead of raw data.However,shared parameters remain vulnerable to inference attacks,e.g.differential attacks,which can still compromise individual privacy.To mitigate such threats,differential privacy(DP)has been widely adopted in FL frameworks.This paper investigates a client-level DP federated learning scenario with noise optimization,and proposes an adaptive DP-FL algorithm based on personalized local momentum.Specifically,we first implement dynamic calibration of local models through client-specific momentum mechanisms to address the client drift problem.Furthermore,we design an adaptive DP mechanism fea-turing decaying clipping thresholds,which dynamically optimizes the noise injection scale for individual clients.Experimental evaluations on Cifar10,Cifar100,and SVHN datasets demonstrate that our algo-rithm outperforms existing approaches under equivalent privacy protection levels.