CSC: Turning the Attacker's Poison Against Their Own Backdoor Poisoning Defense

The study proposes a backdoor poisoning defense method, CSC: it extracts features in early epochs and uses DBSCAN clustering to identify anomalous clusters to isolate poisoned samples, then relabels them as virtual classes and fine-tunes the classifier to replace backdoor associations with benign virtual associations. On 4 benchmark datasets and 12 poisoning attacks, CSC reduces the average attack success rate to near zero compared with 9 SOTA defenses, with little loss in clean accuracy.