/False-Friends

[AAAI 2022] With False Friends Like These, Who Can Notice Mistakes?

Primary LanguagePythonMIT LicenseMIT

With False Friends Like These, Who Can Notice Mistakes?

Code for AAAI 2022 paper "With False Friends Like These, Who Can Notice Mistakes?" by Lue Tao, Lei Feng, Jinfeng Yi, and Songcan Chen. In the paper, we unveil the threat of hypocritical examples---inputs that are originally misclassified yet perturbed by a false friend to force correct predictions. This repository contains an implementation of the attack (hypocritical examples) and several countermeasures (PGD-AT, TRADES, and THRM).

Getting Started

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

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

Running Experiments

1- Naturally train a classifier on CIFAR-10 (the dataset will be automatically download).

python train_cifar10.py --train_loss ST

2- Construct low-quality training data.

python make_data_cifar10.py --data_type Naive
python make_data_cifar10.py --data_type Noise
python make_data_cifar10.py --data_type Mislabeling
python make_data_cifar10.py --data_type Poisoning

3- Naturally train classifiers and evaluate their performance.

python train_cifar10.py --train_loss ST --data_type Naive
python train_cifar10.py --train_loss ST --data_type Noise
python train_cifar10.py --train_loss ST --data_type Mislabeling
python train_cifar10.py --train_loss ST --data_type Poisoning
python train_cifar10.py --train_loss ST --data_type Quality

4- Adversarially train classifiers and evaluate their performance.

python train_cifar10.py --train_loss AT --data_type Poisoning
python train_cifar10.py --train_loss AT --data_type Quality
python train_cifar10.py --train_loss TRADES --data_type Poisoning
python train_cifar10.py --train_loss TRADES --data_type Quality

5- Compare TRADES and THRM.

python tradeoff_cifar10.py --train_loss TRADES --b 1
python tradeoff_cifar10.py --train_loss TRADES --b 5
python tradeoff_cifar10.py --train_loss TRADES --b 10
python tradeoff_cifar10.py --train_loss TRADES --b 20
python tradeoff_cifar10.py --train_loss TRADES --b 40
python tradeoff_cifar10.py --train_loss TRADES --b 60
python tradeoff_cifar10.py --train_loss TRADES --b 80
python tradeoff_cifar10.py --train_loss TRADES --b 100
python tradeoff_cifar10.py --train_loss THRM --b 1
python tradeoff_cifar10.py --train_loss THRM --b 5
python tradeoff_cifar10.py --train_loss THRM --b 10
python tradeoff_cifar10.py --train_loss THRM --b 20
python tradeoff_cifar10.py --train_loss THRM --b 40
python tradeoff_cifar10.py --train_loss THRM --b 60
python tradeoff_cifar10.py --train_loss THRM --b 80
python tradeoff_cifar10.py --train_loss THRM --b 100

Results

Key takeaways:

  • The security risk of false friends is pervasive: we find that many types of substandard models are vulnerable to hypocritical examples across multiple datasets.
  • We tried to circumvent the threat of false friends via several countermeasures. Results demonstrate the effectiveness of the countermeasures, while the risk remains non-negligible even after adaptive robust training.

Citing this work

@inproceedings{tao2022false,
    title={With False Friends Like These, Who Can Notice Mistakes?},
    author={Tao, Lue and Feng, Lei and Yi, Jinfeng and Chen, Songcan},
    booktitle={Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},
    year={2022}
}