Algorithmic Stability Based Generalization Bounds for Adversarial Training

---

Code for the paper Algorithmic Stability Based Generalization Bounds for Adversarial Training

To run the code

Run python train.py to reproduce the experimental results in the paper.

For example, to reproduce the result of tanh_{\gamma}-PGD AT for \gamma = 10 on CIFAR-10, run

python train.py --dataset 'cifar10' --att_method 'tanh'  --beta 10

To reproduce the result of G_{p}-PGD AT for p=2 on CIFAR-10, run

python train.py --dataset 'cifar10' --att_method 'norm_steep'  --beta 2

Experimental results will be saved in the folder cifar10_results, cifar100_results or svhn_results