Impacts of adversarial inputs in associative memory models and its iterative learning variants

Abstract

Adversarial attacks have always been a bane to neural networks. Most of the research focus is on adversarial networks and its defence for deep neural networks. Associative memory models are different class of neural models used in image recognition tasks. There are fundamental differences between Deep neural networks and Associative memory models in terms of the learning procedures. These fundamental differences in turn have different effects on adversarial attacks. In this paper we have attempted an empirical study on various flavors of an associative memory models viz.Hopfield model and two different forms of iterative learning rules and its resilience towards an adversarial attack.