Memristive Crossbar Array-Based Adversarial Defense Using Compression

Abstract

This paper shows that Memristive Crossbar Array (MCA)-based neuromorphic architectures provide a robust defense against adversarial attacks due to the stochastic behavior of memristors. Furthermore, it shows that adversarial robustness can be further improved by compression-based preprocessing steps that can be implemented on MCAs. It also evaluates the effect of inter-chip process variations on adversarial robustness using the proposed MCA implementation and studies the effect of on-chip training. It shows that adversarial attacks do not uniformly affect the classification accuracy of different chips. Experimental evidence using a variety of datasets and attack models supports the impact of MCA-based neuromorphic architectures and compression-based preprocessing implemented using MCA on defending against adversarial attacks. It is also experimentally shown that the on-chip training results in high resiliency to adversarial attacks in all chips.