Identifying and Understanding Cross-Class Features in Adversarial Training (ICML 2025)

This is the official repository for ICML 2025 paper Identifying and Understanding Cross-Class Features in Adversarial Training, including code for adversarial training on CIFAR-10/100, class-wise feature extraction, and class-wise feature attribution correlation matrix visualization.

Requirements

• Python 3.9+
• Common libs like torch, torchvision, numpy, tqdm, matplotlib, pandas.
• Data: CIFAR datasets download to /data/cifar_data by default (edit dataset paths in code if needed).

Usage

• Train (run from repo root)
  • python -m train --fname AT_run --dataset cifar10 --mode AT --norm linf --ne 200 --lr 0.1 --bs 128
  • Modes: AT, TRADES, none
  • Outputs: train_log/<fname>/config.json, log.csv, model_best.pth, model_last.pth, periodic model_<epoch>.pth, opt_<epoch>.pth
• Extract features
  • python -m show_feature --fname AT_run --model train_log/AT_run/model_best.pth --arch PRN18 --dataset cifar10 --norm linf --eps 8
  • Saves: features/<fname>/test_clean.pth and features/<fname>/test_adv.pth
• Correlation and figures
  • python -m evaluate_matrix --fname AT_run --model train_log/AT_run/model_best.pth --arch PRN18 --dataset cifar10 --norm linf --eps 8
  • Saves: correlation matrices and optional diffs to figs/<fname>/
  • Note: you can load features from features/<fname>/test_clean.pth and features/<fname>/test_adv.pth to avoid re-extracting features.

Citation

@inproceedings{wei2025identifying,
  title={Identifying and Understanding Cross-Class Features in Adversarial Training},
  author={Wei, Zeming and Guo, Yiwen and Wang, Yisen},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2025}
}