VGG-face数据集通用对抗模板全NaN问题技术求助

Hey Shashank, sorry to hear you're stuck on this NaN problem with your face classification adversarial attack project—let's walk through some targeted fixes based on common pitfalls when adapting this codebase from ImageNet to VGG-Face:

• Fix Input Normalization Mismatch
  The original code is tuned for ImageNet's normalization rules (mean values [103.939, 116.779, 123.68]), but VGG-Face uses its own standard preprocessing (mean [129.1863, 104.7624, 93.5940]). If your face images aren't normalized correctly for VGG-Face, the model's output scores (ff(idx)) can become numerically unstable, causing project_boundary_polyhedron to fail and return 0.

  • Action: Replace all ImageNet-specific preprocessing code with VGG-Face's required normalization steps. Double-check that your input tensors match the model's expected range.

• Debug ff(idx) and ddf Values
  A return value of 0 from project_boundary_polyhedron usually means it can't find a valid projection direction. This often happens if:

  • ff(idx) (the score of the target class) is almost identical to other class scores, making the classification boundary undefined.
  • ddf (the gradient of the loss w.r.t. input) is all zeros, which breaks the projection calculation.
  • Action: Add debug prints right before calling the function to check the actual values of ff(idx) and the mean/variance of ddf. If ddf is all zeros, ensure your VGG-Face model isn't fully frozen—you need gradients to flow for DeepFool to work.

• Tweak the Polyhedron Projection Parameter Q
  The Q parameter in the original code is set for ImageNet's 1000-class distribution. VGG-Face has 2622 classes, with different class spacing and classification difficulty, so the original Q might be too restrictive to find a valid projection.

  • Action: Try increasing Q by 10-20% (start small to avoid over-perturbing) or calculate a new Q based on VGG-Face's typical class score differences. Q defines the slack for the classification boundary, so it needs to fit your target dataset.

• Prevent Numerical Overflow/Underflow
  Even if you switched image data types, intermediate calculations in project_boundary_polyhedron might suffer from float32 precision loss, leading to NaNs. Division by tiny values (near zero) is another common culprit.

  • Action: Switch all relevant calculations to float64 to boost precision. Add a small epsilon (like 1e-8) to any denominator terms in the projection function to avoid division by zero.

• Ensure Model-DeepFool Compatibility
  The original DeepFool implementation is built for the author's ImageNet model. VGG-Face has a different final fully connected layer dimension, which might cause shape mismatches between ddf and the input expected by project_boundary_polyhedron. Also, make sure your model outputs logits, not softmax probabilities—softmax can flatten gradients and cause numerical issues.

  • Action: Verify that ddf has the correct shape matching the model's output. If you're using a pre-trained VGG-Face model, confirm it's set to output raw logits instead of normalized probabilities.

内容的提问来源于stack exchange，提问作者shawshark