GCBA : Gradient Compensated Backdoor Attack Against Federated Learning

ABSTRACT Federated learning (FL) enables distributed model training while keeping user data private. However, its decentralized structure makes it vulnerable to backdoor attacks. Prior studies show that adversaries can implant backdoors, but these are often weakened by subsequent benign updates, reducing their persistence. In this paper, we propose GCBA, a new backdoor mechanism for FL. GCBA compensates for the average gradient of honest clients, forcing the aggregated update to match the attacker's target. It also restricts perturbations to coordinates with low historical variance, improving durability and stealth. Experiments on three large‐scale datasets show that GCBA achieves higher success rates and longer lifespans than existing attacks. It also evades several representative defenses. These results highlight the urgent need for stronger protection in practical FL systems.