Defending Against Model Poisoning Attacks in Federated Learning via Client-guided Trust

Federated learning (FL) enables multiple data owners to collaboratively train a global model without sharing their private training data. However, FL is threatened by the model poisoning attacks. Most existing defense methods rely on a clean validation dataset on the server to identify attacks, which contradicts the privacy-preserving principle of FL. Furthermore, defending against a substantial proportion of malicious clients remains a significant challenge. This paper proposes CT-BRFL, a Byzantine-robust FL framework where clients guide trust. Experiments show that even one benign client can effectively guide trust. The server starts with an initial trusted set of benign clients and iteratively selects new trusted clients based on guidance from existing ones. At the same time, it detects and removes potential malicious clients to maintain reliability. Trusted clients update the global model using a robust aggregation rule. Extensive experiments are conducted with five aggregation rules under six model poisoning attacks on two datasets. The experimental results demonstrate that on MNIST, CT-BRFL can achieve a detection accuracy of 100% even with more than 50% Byzantine clients, while improving model accuracy by up to 25%.