On Adversarial Training without Perturbing all Examples (SAT)

Official repository for "On Adversarial Training without Perturbing all Examples", Accepted at ICLR 2024.
Paper PDF, Reviews
Poster: poster.pdf

In a Nutshell: Vanilla adversarial training (AT) and most its variants perturb every training example. To what extent is that necessary? We split the training set into subsets A and B, train on AuB but construct adv. examples only for examples in A.

Requirements

• python 3.8
• pytorch 1.6.0
• autoattack
• tensorboard
• apex

Reproducing Results

All our experiments are represented by yaml config files. They can be found in the directory 'config/'. To train on multiple GPUs, make sure to update the value for the config key train_gpu, e.g. train_gpu: [0,1,2,3].

Training

To train, e.g. ESAT on ImageNet-200, run bash config/imagenet200/weightedincreasing2to10_esat_pgd7_decreasing_entropy/run_all.sh. This trains 10 models, according to the config files defined at the same location: config/imagenet200/weightedincreasing2to10_esat_pgd7_decreasing_entropy/*.yaml.

Evaluation

After training completion, adversarial robustness is evaluated via AutoAttack. To start, call bash config/imagenet200/weightedincreasing2to10_esat_pgd7_decreasing_entropy/eval_all.sh

Additional Experiments

Results on Wide-ResNets (discussed here) training script can be found here: 'config/cifar10/wrn70-16_esat_pgd7_decreasing_entropy/'.