A fine tuning technique over the adversarially trained models to increase further robustness
Running the code
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
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
Training via LAT
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.