FLRAM: Robust Aggregation Technique for Defense against Byzantine Poisoning Attacks in Federated Learning
Haitian Chen, Xuebin Chen, Lulu Peng, Ruikui Ma
Abstract
In response to the susceptibility of federated learning, which is based on a distributed training structure, to byzantine poisoning attacks from malicious clients, resulting in issues such as slowed or disrupted model convergence and reduced model accuracy, we propose a robust aggregation technique for defending against byzantine poisoning attacks in federated learning, known as FLRAM. First, we employ isolation forest and an improved density-based clustering algorithm to detect anomalies in the amplitudes and symbols of client local gradients, effectively filtering out gradients with large magnitude and angular deviation variations. Subsequently, we construct a credibility matrix based on the filtered subset of gradients to evaluate the trustworthiness of each local gradient. Using this credibility score, we further select gradients with higher trustworthiness. Finally, we aggregate the filtered gradients to obtain the global gradient, which is then used to update the global model. The experimental findings show that our proposed approach achieves strong defense performance without compromising FedAvg accuracy. Furthermore, it exhibits superior robustness compared to existing solutions.