BAN: Detecting Backdoors Activated by Adversarial Neuron Noise

Environment settings

check requirements.txt

Detection

CUDA_VISIBLE_DEVICES=0 python ban_detection.py --checkpoint ${ckpt} --arch resnet18 --mask-lambda 0.75 --steps 30 --eps 0.3 --print-every 20

Mitigation

CUDA_VISIBLE_DEVICES=0 python ban_finetune.py --epoch 25 --attack_mode all2one --lr 0.005 --rob-lambda 0.5 --checkpoint ${ckpt} --poison-type bpp --trigger-alpha 1.0

Pretrained example models

You can download the trained backdoor models: link