harry24k/ipnn-pytorch

A pytorch implementation of "Intriguing properties of neural networks"

IPNN-pytorch

A pytorch implementation of "Intriguing properties of neural networks"

Summary

1. The natural basis is not better than a random basis for inspecting the properties of latent vectors.
   • there are serveral directions which have the semantic meaning not only individual units.
2. We can generate images(adversarial examples) with small perturbations to fooling neural network models.
3. Weight Decay or Regularization couldn't help model to defend adversarial examples.
4. One adversarial example for a specific model is possible to deceive other models.
5. According to spectral analysis of unstability, the deeper models, the more stupid.

Requirements

• python==3.6
• numpy==1.14.2
• pytorch==1.0.0

A limitation of this code

In the paper, L-BFGS is used to solve equation with constraints.
However, in this code, backpropagation method is used instead of L-BFGS.
Hence, it doesn't cover "4.2 Exprimental results"

Notice

• This Repository won't be updated.
• Please check the package of adversarial attacks in pytorch