Backdoor Evaluation
AFU-IC's backdoor accuracy closely tracked retraining over 10 post-learning rounds on all three datasets. Projected-GA and FedRecovery initially lowered backdoor accuracy but reverted strongly after continued training. Lower backdoor accur…
1 sources - 4 claims
AFU-IC's backdoor accuracy closely tracked retraining over 10 post-learning rounds on all three datasets. Projected-GA and FedRecovery initially lowered backdoor accuracy but reverted strongly after continued training. Lower backdoor accuracy indicates better removal of backdoor behavior. Unlearning completeness was evaluated with a backdoor proxy where the target client poisons part of its local data.