search

hcguoO0 / FIA

code for "Feature Importance-aware Transferable Adversarial Attacks"
74 stars 15 forks source link

Parameter setting of resnet_v1_152 #5

Closed wowowoxuan closed 2 years ago

wowowoxuan commented 2 years ago

Hi, thank you for sharing the code! I'm trying to run the attack.py using resnet_v1_152, may I know the parameter setting.

hcguoO0 commented 2 years ago

As written in our paper, the drop probability pd = 0.3 (probb=0.7) when attacking normally trained models and pd = 0.1 (probb=0.9) when attacking defense models, and the ensemble number N = 30 (ens=30) in aggregate gradient, and the layer parameter is the last layer of second block for Res-152 (layer_name=resnet_v1_152/block2/unit_8/bottleneck_v1/add)

wowowoxuan commented 2 years ago

Got it, thanks a lot!