Probabilistic Label Flipping Attack in Federated Learning

Federated Learning (FL) is vulnerable to adversarial attacks due to its decentralized nature, with label-flipping attacks being a major threat. Unlike prior studies that assume static attacker behavior, this work introduces probabilistic label-flipping attacks, where adversaries participate stochastically, leading to a varying number of attackers across rounds instead of remaining fixed. Attackers decide whether to launch label-flipping attacks in a given round based on a probability α. We evaluate the impact of attack probability (α =0.5, 0.7, 1.0) and attacker percentage (π = 10%, 30%) on FL performance. The results show that higher attack probability and increased attacker participation cause severe performance degradation, with the most significant drop at π = 30%, α = 1.0. Performance metrics of both local and global models deteriorate significantly, emphasizing the need for adaptive defense mechanisms against intermittent adversarial behavior in FL.