一种抗拜占庭攻击的联邦学习鲁棒聚合算法OA

To address the issue that existing defense schemes in federated learning tend to over-prune benign mod-els during filtering,a robust aggregation algorithm defending against Byzantine attacks in federated learning(FLD-BA)was proposed.HDBSCAN density-based clustering was employed to group models,to identify the benign clus-ter,and the most representative model in terms of direction was selected as the trusted reference model.Using the trusted model as a benchmark,cosine similarity was utilized to screen potentially misclassified benign models within clusters,thereby to correct misjudgments.Additionally,a reputation mechanism was established to dynamically e-valuate models' historical behaviors,to mitigate the impact of missed detections.For models with high reputation,adaptive magnitude scaling was applied,and differential aggregation weights were assigned based on update quality to further enhance aggregation performance.Experimental results demonstrated that when defending against sign-flipping attacks,FLDBA achieved an accuracy improvement of 0.18 percentage points to 5.13 percentage points compared to FLRAM,FLAME,RFLPA,FLTrust,and Krum,while reducing the attack success rate by 40.52 per-centage points to 61.39 percentage points,exhibiting superior robustness.