The author introduces the concept of backdoor-critical layers in federated learning, proposing new attack methodologies that exploit these layers for stealthy attacks.
Backdoor attacks can manipulate agents without affecting performance, prompting the need for robust defense mechanisms.
The author proposes a novel backdoor attack mechanism, DPOT, that effectively conceals malicious clients' model updates among those of benign clients by dynamically adjusting backdoor objectives, rendering existing defenses ineffective.