Robustness of Sparsely Distributed Representations to Adversarial Attacks in Deep Neural Networks

Abstract

Deep learning models have achieved an impressive performance in a variety of tasks, but they often suffer from overfitting and are vulnerable to adversarial attacks. Previous research has shown that dropout regularization is an effective technique that can improve model generalization and robustness. In this study, we investigate the impact of dropout regularization on the ability of neural networks to withstand adversarial attacks, as well as the degree of “functional smearing” between individual neurons in the network. Functional smearing in this context describes the phenomenon that a neuron or hidden state is involved in multiple functions at the same time. Our findings confirm that dropout regularization can enhance a network’s resistance to adversarial attacks, and this effect is only observable within a specific range of dropout probabilities. Furthermore, our study reveals that dropout regularization significantly increases the distribution of functional smearing across a wide range of dropout rates. However, it is the fraction of networks with lower levels of functional smearing that exhibit greater resilience against adversarial attacks. This suggests that, even though dropout improves robustness to fooling, one should instead try to decrease functional smearing.