LAT adversarial_robustness

A fine tuning technique over the adversarially trained models to further increase adversarial robustness.

Code for the paper Harnessing the Vulnerability of Latent Layers in Adversarially Trained Models : https://arxiv.org/abs/1905.05186

Running the code

Dataset: CIFAR10

Fetching LAT robust model

The model can be downloaded from this link - https://drive.google.com/open?id=1um2zoVYYw5YZuuV8_IeoUy-qRWSmCVUb> .

Evaluating the LAT robust model

python eval.py

The trained model achieves test accuracy of 87.8% and adversarial robustness of 53.82% against PGD attack(epsilon = 8.0/255.0)

Fetching Adversarial Trained Model

python fetch_model.py

Training via LAT

python feature_adv_training_layer11.py

Latent Attack

python latent_adversarial_attack.py

Example original and adversarial images computed via Latent Attack on CIFAR10

Example original and adversarial images computed via Latent Attack on Restricted Imagenet(https://arxiv.org/pdf/1805.12152.pdf <https://arxiv.org/pdf/1805.12152.pdf>).

Citing this work

@article{nupur2019lat,
  title={Harnessing the Vulnerability of Latent Layers in Adversarially Trained Models},
  author={Kumari, Nupur and Singh, Mayank and Sinha, Abhishek and  Krishnamurthy, Balaji and  Machiraju,Harshitha and   Balasubramanian, Vineeth N},
  journal={IJCAI},
  year={2019}
}