search

paarthneekhara / AdversarialDeepFakes

27 stars 5 forks source link

MesoNet cannot Classify Original Videos accurately #17

Open kavilan-nair opened 2 years ago

kavilan-nair commented 2 years ago

I am trying to test different generative deepfake defense mechanisms on both Xception and MesoNet models.

When I run the detect_from_video.py script using the linked pretrained mesonet weights, I am seeing that it classifies most original images as deepfakes. I am not sure if I am doing something wrong when loading the model in to run inference. Could you possibly confirm if this is the same behaviour you see.

I noticed that your paper didn't test the adversarial attacks on the original class too. Was there a reason for this?

BenPinhasov commented 1 year ago

@kavilan-nair I have this issue too. did you solve it?