REDFL: Robust and Efficient Decentralized Federated Learning Based on Committee Consensus

Federated learning (FL) enables multiple clients to collaboratively train a global model without disclosing their original data. However, traditional FL is susceptible to attacks from malicious clients who can disrupt the training process by uploading poisoned models to the server. In this paper, we introduce REDFL, a robust and efficient decentralized federated learning framework that employs committee consensus to effectively defend against poisoning attacks. Our framework not only enhances security but also improves the performance of the global model. In REDFL, we propose a dual model validation algorithm based on mutual information for precise validation of model updates. To evaluate the contributions made by clients to the global model, we develop a contribution calculation method that accounts for the varying roles of clients. Our proof of contribution mechanism enables a more accurate selection of honest clients as committee members. The committee validates the model updates from trainers, reaches a consensus on these updates, and ultimately aggregates a high-quality global model. Experimental results demonstrate that REDFL outperforms other Byzantine-tolerant algorithms in terms of global model accuracy and convergence speed. On the Non-IID dataset under a poisoning attack scenario, REDFL achieves an accuracy of 80%, showcasing its robustness and effectiveness.