FedTruth: A General Scheme for Defending Against Byzantine and Backdoor Attacks in Federated Learning

FedTruth is a general scheme for defending against Byzantine attacks and backdoor attacks in federated learning, estimating the true global model update from all model updates through dynamic aggregation weights. The method requires no server-side benign root dataset, makes no assumptions about benign or malicious data distributions, and considers the contributions of all benign clients. Experiments show that FedTruth can reduce the influence of poisoned model updates and remain efficient in large-scale federated learning systems.