Adversarial training with Advprop improves accuracy but decreases robustness

Colab Notebook for the Advanced Computer Vision project focused on experiments with the AdvProp framework for adversarial training of DNNs

Reference paper

The work is based on the paper "Adversarial Examples Improve Image Recognition" by Xie et al. (CVPR 2020): https://arxiv.org/abs/1911.09665.

Overview

We use the EfficientNet implementation of https://github.com/aladdinpersson/Machine-Learning-Collection/blob/master/ML/Pytorch/CNN_architectures/pytorch_efficientnet.py on top of which we built experiments with different Batch Norm layers, Domain Adaptation techniques and second-order adversarial attacks for testing both accuracy and robustness of the deep neural networks.