Hi,
I was trying to generate adversarial examples using Jacobian based saliency map method. I was using your implementation. First of all, I would like to know whether your implementation of JSMA is correct or not. Have you tested it? I just want to make sure that it is flawless.

I am using the pre-trained inceptionv3 model on pytorch and I would like to know what are the expected dimensions of inputs of jsma function. I pass 1X3X299X299 dimensional matrix as an input_tensor. I compute target_class as follows: y_target = torch.LongTensor([9]). Then I call jsma function as follows: jsma(inceptionv3, image_tensor, y_target, max_distortion=0.1). It runs successfully but when I compare output with image_tensor-both are same. This is not an expected output. Can you please guide me here? what went wrong?

Also, note that here input_features.size(1) will return 3 and it will be multiplied by 0.1 to calculate num_features, this will be 0.3. Therefore, only one iteration will be performed.
`