This repo implements two adversarial training algorithms:

PGD-AT: Towards Deep Learning Models Resistant to Adversarial Attacks

TRADES: Theoretically principled trade-off between robustness and accuracy

To run the model, simply run the following command:

python algorithms/trainer_trades.py

with settings specified in trainer_at.py or trainer_trades.py.

The checkpoints are also given in trained_model with best model named as best.pth.

In PGD-AT, the initial point is randomly generated in the tolerable attack range. In TRADES, the initial point is randomly generated in a small range according to the original paper.

We train each robust model with 150 epochs, though in most of the case, 100 epochs are enough to have a good result.

The robust models are evaluated using foolbox