README.md

Recommendation: We suggest using the unified implementation in our open-source FastAT Benchmark for easier reproduction and fair comparison.

This repository contains the official PyTorch implementation for our ICCV 2025 paper: Mitigating Catastrophic Overfitting in Fast Adversarial Training via Label Information Elimination.

Setup

1. Clone the repository

git clone https://github.com/fzjcdt/LIET.git
cd LIET

2. Install dependencies

pip install -r requirements.txt

Environment

This code has been tested on the following environment:

• OS: Ubuntu 20.04.3
• GPU: Tesla V100
• CUDA: 11.4
• Python: 3.8.10
• PyTorch: 1.10.1
• Torchvision: 0.11.2

Usage

Training

To start training the model from scratch, run the main script:

./main.sh

This script will handle the entire training and testing processes as described in the paper.

Reproducing Figure 1, Figure 2 and Table 1

To reproduce Figure 1, Figure 2, and Table 1 from our paper, please refer to the scripts and instructions provided in the ./paper_figures directory.

Citation

If you find this work useful for your research, please consider citing our paper:

@inproceedings{pan2025mitigating,
  title={Mitigating Catastrophic Overfitting in Fast Adversarial Training via Label Information Elimination},
  author={Pan, Chao and Tang, Ke and Li, Qing and Yao, Xin},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={2991--3000},
  year={2025}
}