Dual Filtering Based Decentralized Learning Against Byzantine Attack

Decentralized deep learning improves upon centralized methods by reducing failure risks but faces challenges from Byzantine attacks that can disrupt model integrity. In Byzantine defense models, the main approach is to rank neighboring node scores, which places strict demands on network topology.This paper proposes a Dual Filtering based Parameter Aggregation Rule (DF-PAR). It primarily focuses on the parameter cosine similarity and distance, employing predefined thresholds to filter out the outliers. Additionally, an evaluation mechanism rooted in the credibility of historical information is introduced to determine the weight value of each node. Eventually, the impact of the suspicious node is gradually reduced and eliminated. The experimental results demonstrate the algorithm’s ability and convergence even when some benign nodes are surrounded by Byzantine nodes.