FLEvaluate: Robust Federated Learning Based on Trust Evaluate
Chao Guo, Buxin Guo, Tingting Zhu, Peihe Liu, Gong Cheng · 2023
Federated Learning (FL) has emerged as a popular distributed learning framework, enabling participants to collaboratively train models without the need to share raw data. In contrast to traditional machine learning frameworks that require the collection of large volumes of user data for centralized training, FL provides a novel approach that ensures data privacy and addresses the issue of data silos. However, the conventional FL framework faces challenges posed by the Byzantine Generals' Problem, where malicious clients can easily compromise the global learning process. Existing Byzantine solutions in FL are ineffective in defending against such attacks. This paper proposes a novel FL method called FLEvaluate, which assesses client trust based on performance metrics of their models and assigns aggregation weights according to their trust levels. By incorporating trust evaluation and normalization of consequences, this method amplifies or reduces the impact of clients in the global model aggregation process, thereby achieving a robust FL framework. Comparing FLEvaluate with the traditional FL framework FedAvg reveals that not only does FLEvaluate achieve comparable accuracy and precision but it also demonstrates effective defense against Byzantine attacks.