Multimodal Feature Fusion Against Poisoning Attacks in Federated Graph Learning

Recently, Federated Graph Learning (FedGL) has been widely used in personalized healthcare recommendations due to its strong privacy protection and collaborative learning across entities. However, the separation between federated clients and the central server makes FedGL vulnerable to model poisoning attacks. These attacks can spread quickly through the graph structure and are further obscured by client privacy mechanisms. Existing defense methods often fail to contain the spread of attacks and lack adaptive attacker identification. To address these challenges, we propose TS-Defense, a feature fusion framework based on the Transformer and Statistical techniques. TS-Defense enables robust detection and purification of malicious clients in FedGL. It introduces a two stage defense approach: (1) a balanced graph partitioning approach to isolate subdomains and limit the spread of poisoned gradients; (2) a multimodal detection and forgetting module for accurate identification and purification of malicious gradients. The efficiency and effectiveness of TS-Defense are comprehensively evaluated. Experimental results show that TS-Defense can effectively mitigate poisoning attacks, has minimal impact on the aggregated model’s performance, maintains stable detection efficiency, and achieves desirable time overhead.