Code for NeurIPS 2021 paper "Better Safe Than Sorry: Preventing Delusive Adversaries with Adversarial Training" by Lue Tao, Lei Feng, Jinfeng Yi, Sheng-Jun Huang, and Songcan Chen.
This repository contains an implementation of the attacks (P1~P5) and the defense (adversarial training) in the paper.

Our code relies on PyTorch, which will be automatically installed when you follow the instructions below.

conda create -n delusion python=3.8
conda activate delusion
pip install -r requirements.txt

1. Pre-train a standard model on CIFAR-10 (the dataset will be automatically download).

python main.py --train_loss ST

2. Generate perturbed training data.

python poison.py --poison_type P1
python poison.py --poison_type P2
python poison.py --poison_type P3
python poison.py --poison_type P4
python poison.py --poison_type P5

3. Visualize the perturbed training data (optional).

tensorboard --logdir ./results

4. Standard training on the perturbed data.

python main.py --train_loss ST --poison_type P1
python main.py --train_loss ST --poison_type P2
python main.py --train_loss ST --poison_type P3
python main.py --train_loss ST --poison_type P4
python main.py --train_loss ST --poison_type P5

5. Adversarial training on the perturbed data.

python main.py --train_loss AT --poison_type P1
python main.py --train_loss AT --poison_type P2
python main.py --train_loss AT --poison_type P3
python main.py --train_loss AT --poison_type P4
python main.py --train_loss AT --poison_type P5

Table 1: Below we report mean and standard deviation of the test accuracy for the CIFAR-10 dataset. As we can see, the performance deviations of the defense (i.e., adversarial training) are very small (< 0.50%), which hardly effect the results. In contrast, the results of standard training are relatively unstable.

Key takeaways: Our theoretical justifications in the paper, along with the empirical results, suggest that adversarial training is a principled and promising defense against delusive attacks.

@inproceedings{tao2021better,
    title={Better Safe Than Sorry: Preventing Delusive Adversaries with Adversarial Training},
    author={Tao, Lue and Feng, Lei and Yi, Jinfeng and Huang, Sheng-Jun and Chen, Songcan},
    booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
    year={2021}
}