Sybil attacks detection for dynamic environment in federated learning

Abstract

Federated learning can utilize its distributed structure to protect data privacy security of clients and improve efficiency of machine learning. However, its distributed framework also make itself be susceptible to sybil attacks. While previous research has already proposed defense methods to address this issue, they often fail to guarantee effective performance in a dynamic federated learning system, where some clients dynamically join in and out. To tackle this problem, our paper introduces a novel defense method specifically designed to mitigate sybil attacks in dynamic federated learning scenario. Our proposed method consists of three mechanisms: similarity mechanism, validation mechanism, and reputation mechanism. These mechanisms can address the problem of missing information and effectively resist sybil attacks in dynamic federated learning. We evaluate the performance of our method on the MNIST and KDDCup datasets and demonstrate its advanced ability in defending against sybil attacks in dynamic federated learning compared to existing methods.