Instance Adaptive Adversarial Training

Code accompanying out paper: "Instance adaptive adversarial training: Improved accuracy tradeoffs in neural nets"

Requirements

• Pytorch
• Foolbox
• Adversarial robustness toolbox

Training commands

Please update config json files to modily parameters/ specify data paths.

Training a clean model (ERM) -

python main.py --cfg-path configs/train.json --data_root [specify path] --alg clean 

Training adversarial training -

python main.py --cfg-path configs/train.json --data_root [specify path] --alg adv_training 

Training instance adaptive algorithm (our approach)

python main.py --cfg-path configs/train_adaptive.json --data_root [specify path]

To evaluate a model, run

python main.py --cfg-path configs/eval.json --data_root [specify path] --save_path [path to log results] --attack_steps [number of attack steps] --restore [path to restore saved model from]