Implicit adversarial data augmentation and robustness with Noise-based Learning

We introduce a Noise-based Learning (NoL) approach for training neural networks that are intrinsically robust to adversarial attacks. find the learning of random noise introduced with input same loss function used during posterior maximization, improves model’s resistance. show learnt performs implicit data augmentation boosting adversary generalization capability. evaluate our approach’s efficacy and provide simplistic visualization tool understanding data, using Principal Component Analysis. conduct comprehensive experiments on prevailing benchmarks such as MNIST, CIFAR10, CIFAR100, Tiny ImageNet remarkably well against wide range Furthermore, combining NoL state-of-the-art defense mechanisms, training, consistently outperforms prior techniques in both white-box black-box