Introducing Diversity In Feature Scatter Adversarial Training Via Synthesis

In an attempt to understand how deep learning models interpret inputs, it has been found that they change their prediction when a carefully optimized imperceptible noise termed adversarial perturbation is added to the input data. Many researchers are focusing on developing methods to counter such effects, but such methods do not generalize well to adversarial test data. Recently, Feature-Scatter adversarial training has come up to solve such a problem, but this method uses the traditional adversarial training framework as its basis that cannot generate diverse perturbations.In this paper, we propose an approach that combines both the Feature-Scatter adversarial training and the generator-based adversarial training framework to optimally explore the adversarial data manifold achieving better robust generalization. We perform extensive experimentation across a wide variety of datasets such as Cifar10, Cifar100, and SVHN. Our framework significantly outperforms the state-of-the-art methods against both strong white-box attacks and black-box attacks.