search

392781 / FaceOff

Steps towards physical adversarial attacks on facial recognition
Creative Commons Attribution Share Alike 4.0 International
78 stars 14 forks source link

Why the Dist. to Target grows with the epoch? #12

Open iDTer opened 2 years ago

iDTer commented 2 years ago

Epoch 28: Loss = -0.5512599 Dist. to Image = 0.9096553 Dist. to Target = 0.3583954 Face detection = True Epoch 29: Loss = -0.5638140 Dist. to Image = 0.9396414 Dist. to Target = 0.3758273 Face detection = True Epoch 30: Loss = -0.5764654 Dist. to Image = 0.9016970 Dist. to Target = 0.3252316 Face detection = True

heath-halle

I feel like the end result is blending the perturbation on the original face image instead of masking it directly. And I found that as the epoch increases, the perturbations appear more and more sparse.

392781 commented 1 year ago

Yeah, this is an issue that results from RGB values being mapped between [0,1]. I think it's some issue regarding allowing an alpha to be applied to RGB values or something of that sort. I am not actively developing this so if you pin-point the specific problem, feel free to either comment here or make your own pull-request.

392781 commented 1 year ago

I am happy to bounce ideas back and forth with you too!