Neural networks are vulnerable to adversarial attacks -- small visually imperceptible crafted noise which when added to the input drastically changes the output. The most effective method of defending against these adversarial attacks is to use the methodology of adversarial training. We analyze the adversarially trained robust models to study their vulnerability against adversarial attacks at the level of the latent layers. Our analysis reveals that contrary to the input layer which is robust to adversarial attack, the latent layer of these robust models are highly susceptible to adversarial perturbations of small magnitude. Leveraging this information, we introduce a new technique Latent Adversarial Training (LAT) which comprises of fine-tuning the adversarially trained models to ensure the robustness at the feature layers. We also propose Latent Attack (LA), a novel algorithm for construction of adversarial examples. LAT results in minor improvement in test accuracy and leads to a state-of-the-art adversarial accuracy against the universal first-order adversarial PGD attack which is shown for the MNIST, CIFAR-10, CIFAR-100 datasets.

Recently, Neural networks have seen a huge surge in its adoption due to their ability to provide high accuracy on various tasks. On the other hand, the existence of adversarial examples have raised suspicions regarding the generalization capabilities of neural networks. In this work, we focus on the weight matrix learnt by the neural networks and hypothesize that ill conditioned weight matrix is one of the contributing factors in neural network's susceptibility towards adversarial examples. For ensuring that the learnt weight matrix's condition number remains sufficiently low, we suggest using orthogonal regularizer. We show that this indeed helps in increasing the adversarial accuracy on MNIST and F-MNIST datasets.