Fed-BRMC: Byzantine-Robust Federated Learning via Dual-Model Contrastive Detection

Federated Learning (FL) is a distributed privacy-protecting machine learning paradigm that enables collaborative training among multiple parties without the need to share raw data. This mode of training renders FL particularly susceptible to attacks. A pivotal challenge within FL is the defense against Byzantine attacks, which can alter the model aggregation process and ultimately lead to the failure of Federated Learning convergence. Existing Byzantine defense mechanisms often result in diminished model accuracy in scenarios characterized by non-independent and identically distributed (non-IID) data. To alleviate this issue, we propose a novel defense approach named Fed-BRMC. This method combines a self-developed federated Dual-Model Contrastive Detection technique with Similarity Detection methods, leveraging historical update information from users to deeply mine the characteristics of model parameters, effectively distinguishing between Byzantine and honest users. Extensive experimentation has shown that Fed-BRMC significantly improves user identification accuracy in non-IID data scenarios compared to existing methods and notably enhances the global model accuracy of Byzantine-robust Federated Learning.