AIDFL: An Information-Driven Anomaly Detector for Data Poisoning in Decentralized Federated Learning

Decentralized Federated Learning eliminates central servers by enabling direct communication among clients. However, this structure introduces significant security challenges, as each client has access to the model parameters. Existing defense mechanisms face significantly reduced effectiveness under non-IID data distributions. To address these challenges, AIDFL is proposed to utilize conditional entropy and mutual information, which are independent of data distribution to detect and mitigate data poisoning attacks in DFL environments. Experimental results demonstrate that AIDFL achieves superior defense under non-IID settings under different poisoning configurations. In particular, this study not only enhances the robustness of DFL but also highlights the critical need for further research on advanced defense strategies against model poisoning attacks in decentralized frameworks. This work serves as a foundation for future exploration of secure DFL systems.