ppwwyyxx / adversarial-face-attack Goto Github PK

My environment is tensorflow 1.8.0, python 2.7.16. I run this code and get an adversarial image.

The adversarial image can successfully fool face classifier, but I found that the perturbation of some points in this adversarial image is larger than eps, which I set as 16.
I hope you can verify this issue.

Can you explain why you use noise = 0.9 * grad + noise instead of grad = 0.9 * grad + noise in line 94 of face_attack.py?

Thank you!

按照教程配置了环境出现问题

Traceback (most recent call last):
  File "face_attack.py", line 216, in <module>
    victim = model.compute_victim(args.data, args.target)
  File "face_attack.py", line 44, in compute_victim
    embeddings = self.eval_embeddings(image_batch)
  File "face_attack.py", line 118, in eval_embeddings
    return self.sess.run(self.embeddings, feed_dict={self.image_batch: batch_arr})
  File "D:\ProgramData\Anaconda3\lib\site-packages\tensorflow\python\client\session.py", line 905, in run
    run_metadata_ptr)
  File "D:\ProgramData\Anaconda3\lib\site-packages\tensorflow\python\client\session.py", line 1116, in _run
    str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (42, 250, 250, 3) for Tensor 'images:0', which has shape '(?, 160, 160, 3)'

请问怎么处理

dist = np.dot(emb, self.victim_embeddings.T).flatten()

how to understand the dist line, it's a somewhat formula?And whether I got wrong in the following steps ?:
I change the dist to calculate cosine similarity dist, but I found most dist of the original images results is close to 0, and most of most is lower than 0( I think I got wrong with this,normal image pair's cosine similarity is [0,0.5],and the dist result of adversarial examples is close to 0.8 or higher

Thank you.

If you don't add a LICENCE, you reserve all the rights in this code. But it's inconvenient to others who want to improve or use your code.

In the face_attack.py, there is 2 line of code:

import lfw as lfw
import align.detect_face as FaceDet

But I don't know where is the lfw and align package.
Could you explain where is the lfw and align package in face_attack.py?