On the Resiliency of an Analog Memristive Architecture against Adversarial Attacks

Bijay Raj Paudel1, Vasileios Pentsos2, Spyros Tragoudas1
1Southern Illinois University Carbondale, 2Southen Illinois University Carbondale


This paper studies the effects of adversarial attacks on Deep Neural Networks (DNN) constructed using memristor crossbar arrays and analog components. The presented experiments show that the analog architecture is more resilient against adversarial attacks than the software implementation. Its resiliency also compares favorably to a recently presented hybrid architecture that also uses the memristor crossbar arrays. The relationship between adversarial resiliency, non-idealities in crossbars, analog implementation of activation functions, and supporting circuits in the analog architecture is evaluated.