Delay model and machine learning exploration of a hardware-embedded delay PUF

Abstract

A special class of Physically Unclonable Functions (PUF) called strong PUFs are characterized as having an exponentially large challenge-response pair (CRP) space. However, model-building attacks with machine learning algorithms have shown that the CRP space of most strong PUFs can be predicted using a relatively small subset of training samples. In this paper, we investigate the delay model of the Hardware-Embedded deLay PUF (HELP) and apply machine learning algorithms to determine its resilience to model-building attacks. The delay model for HELP possesses significant differences when compared with other delay-based PUFs such as the Arbiter PUF, particularly with respect to the composition of the paths which are tested to generate response bits. We show that the complexity of the delay model in combination with a set of delay post processing operations carried out within the HELP algorithm significantly reduce the effectiveness of model-building attacks.