More Explaination of Attack for Face Verification

Threshold

Use '--threshold' to set the threshold of face verification model.
For face-verfication model resnet101sofxmax, it has threshold 1.244, 1.048, 0.597 for FPR 1e-3, 3e-4, 1e-4 respectively.

Attribute

The default setting will generate adversarial examples for all attributes.

If you only want to use one of them, just set '--adv_attribute'. It should be one of ['Blond_Hair', 'Wavy_Hair', 'Young', 'Eyeglasses', 'Heavy_Makeup', 'Rosy_Cheeks', 'Chubby', 'Mouth_Slightly_Open', 'Bushy_Eyebrows', 'Wearing_Lipstick', 'Smiling', 'Arched_Eyebrows', 'Bangs', 'Wearing_Earrings', 'Bags_Under_Eyes', 'Receding_Hairline', 'Pale_Skin', 'all']

Results

All results will be stored into the folder '--save_path'.

It contains i_k_original_img.png(original image), i_k_target_img.png(target image) and i_k_True/False_adv_G(X,cj).png(adversarial image).
i: original image id,
k: target image id,
j: attribute index

Each attack record is written into 'record.txt'.

All options are in the 'verification_attack.py'.