SCRUB-FL: Sanitizing and Cleansing Representations via Unlearning of Backdoors

SCRUB-FL introduces a novel two-phase solution for post-training backdoor removal in Federated Learning (FL), addressing vulnerabilities from malicious clients. It employs spectral analysis and activation clustering to identify suspicious samples, utilizing lightweight Wasserstein Generative Adversarial Networks (WGAN-GP) for trigger pattern representation. Experimental results demonstrate that SCRUB-FL effectively reduces backdoor attack success rates to 3.88% while preserving over 91% accuracy on normal tasks, making it a significant advancement for practitioners concerned with security in decentralized model training.