H. Wang, X. Wu, Z. Huang, and E. P. Xing. "High frequency component helps explain the generalization of convolutional neural networks." CVPR 2020 (Oral).

[Slides] and [Poster]

Highlights

Fig. 1: The central hypothesis of our paper: within a data collection, there are correlations between the highfrequency components and the “semantic” component of the images. As a result, the model will perceive both high-frequency components as well as the “semantic” ones, leading to generalization behaviors counterintuitive to human (e.g., adversarial examples).

Fig. 2: Eight testing samples selected from CIFAR10 that help explain that CNN can capture the high-frequency image: the model (ResNet18) correctly predicts the original image (1st column in each panel) and the high frequency reconstructed image (3rd column in each panel), but incorrectly predict the low-frequency reconstructed image (2nd column in each panel). The prediction confidences are also shown. Details are in the paper.

1. Trade-off between accuracy and robustness (Section 3)
2. Rethinking data before rethinking generalization (Section 4)
3. Re-evaluate the heuristics (BatchNorm seems to promote high-frequency information) (Section 5)
4. Adversarially robust models tend to filter out high-frequency components (Section 6)
5. Similar phenomena are observed beyond image classification (Section 7)

Code Structures

• scripts/resnet.py
  • Pipeline to replicate our results on CIFAR10 data
• utility/
  • attackHelper.py
    • main methods used to generate adversarial examples
  • dataLoader.py
    • load data
  • frequencyHelper.py
    • generate low and high frequency data
  • pdg_attack.py
    • helper for adversarial training

Before using the code

• Install the main dependency: TensorFlow 1.x and Foolbox
• Read the main script. We recommend users to skim through the script befoure usage. We use "todo" to highlight the parts that may require attention.
• Generate data: one can generate the data with utility/frequencyHelper.py

PyTorch Implementation

The results were generated by the TensorFlow code as shared here, but for the friends who prefer PyTorch, Xindi has nicely created some codes to help you start: PyTorch Implementation

Contact

HaohanWang · Xindi Wu · Zeyi Huang