TOFD: Target-Oriented Feature Decoupling against Poisoning Attacks in Split Federated Learning
The episode reviews TOFD, a defense against poisoning attacks in split federated learning. It detects attacks via smashed data, purifies samples, and uses a decoy model to suppress residual effects. Hosts highlight its low overhead, robustness to non-IID data, and theoretical guarantees, noting it outperforms baselines across datasets.