shivangi-aneja / tafim Goto Github PK

Hello @shivangi-aneja, I have a question about the perturbation loss used for the single VS multiple manipulation methods.

In eq.14, you try to minimize $||\delta_i^{all}||_2$ but in eq.6, you try to minimize $||X_i^p-X_i||_2 + ||X_i^{Gp}-X_i||_2$. I understand that both formula are approximately the same since we approximately have $$||X_i^p-X_i||_2=||Clamp_{\epsilon}(X_i+\delta_i)-X_i||_2\approx||\delta_i||_2$$
If my above supposition is correct, why would you use $||\delta_i^{all}||_2$ in eq.14 instead of $||X_i^{all}-X_i||_2$ ?

I suppose that this is not a really important question since there wouldn't be much difference in the result, but still I'd want to know if there is a particular reason for doing that.

In any cases, thank you for this great work. I've learned a lot thanks to your paper.

Hi aneja, Thanks for your amazing contribution. I would like to perform an attack on SimSwap, but I am having some problems with the format of the Arcface checkpoint provided by the SimSwap author as a tar file. Could you kindly provide me with the arcface.pth file as shown in your project configuration. Thanks a lot!

Congratulation for your great work! May I know if you can release the pretrained model of TAFIM? It would be really helpful.

Hi aneja, I trained the Face Protection model together with the SimSwap model under the configuration like this:
net_noise = 'unet_64' (default) n_epochs = 100 perturb_wt=10 (default) lr = 0.0001 (default)
,due to the limitation of computing resources, I only used 2000 images from the CelebHQ dataset and trained the model for 100 epochs. The test results are shown below.

The second colum is the source image with pertubation and the noise seems quite obvious. I am wondering if this means I need to train for more epochs?
And the testing script I used is here: https://github.com/wkyhit/TAFIM/blob/main/testing_scripts/test_protection_model_SimSwap.py