A Weighted Asynchronous Federated Learning Scheme to Combat Model Poisoning Attacks

Résumé fourni par la source

With increasing demands for data privacy protection, federated learning, as a distributed collaborative training mechanism, achieves joint optimization of multi-party models without sharing original data. However, traditional synchronous federated learning suffers from low communication efficiency and high resource consumption due to the need to wait for all clients to complete training before aggregation, making it difficult to adapt to heterogeneous and dynamic network environments. To address this, asynchronous federated learning has been proposed, allowing the server to perform global updates immediately after receiving updates from some clients, significantly improving training efficiency and scalability. However, during asynchronous processes, malicious nodes may inject forged or tampered gradient updates, severely impacting the model’s convergence and reliability. This paper proposes an asynchronous robust weighted federated learning scheme to achieve efficient defense against poisoning attacks in asynchronous environments. The proposed solution integrates Gaussian mixture modeling and Mahalanobis distance for statistical detection, and uses homomorphic encryption to securely collect userencrypted models, ensuring both security and efficiency during model updates. Theoretical analysis shows that this solution is more resistant to model poisoning attacks in asynchronous federated learning scenarios, effectively improving the model’s convergence speed and robustness.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.