search

suzhigangssz / AVIH

Code for Visual Information Hiding Based on Obfuscating Adversarial Perturbations
12 stars 2 forks source link

some doubts #6

Open dreamtojava opened 6 months ago

dreamtojava commented 6 months ago

Thank you very much for your response. I have some more questions, and I look forward to your reply. Thank you very much. 1.Why can only the protected images generated by one's own key model be accurately restored, while other models (even with the same architecture) find it difficult to restore them well? 2.Why can the service model use the image to correctly extract features? Furthermore, these protected images cannot be recognized by other models.

suzhigangssz commented 6 months ago

Thank you very much for your response. I have some more questions, and I look forward to your reply. Thank you very much. 1.Why can only the protected images generated by one's own key model be accurately restored, while other models (even with the same architecture) find it difficult to restore them well? 2.Why can the service model use the image to correctly extract features? Furthermore, these protected images cannot be recognized by other models.

  1. See the end of the section 'Key model randomness analysis.' in section 4.3 of the paper.
  2. We keep the features of the original image and the protected image being extracted by the service model consistent by the loss $L{v}$ ( Eq. 4 of the article).$L{V}$ is mainly computed by a specific service model so that the resulting protected image is not recognizable by other models.