Evaluating the Unpredictability of Multi-Bit Strong PUF Classes

Abstract

With advances in machine learning (ML), cybersecurity solutions and security primitives are becoming increasingly vulnerable to successful attacks. Strong Physical Unclonable Functions (PUFs) emerge as a potential countermeasure that offers high resistance to such attacks. In this paper, we introduce a generalized attack model that leverages the collective responses of multiple PUF chips within the same class to improve the prediction accuracy of responses for unobserved challenges, in contrast to traditional single-chip approaches. Furthermore, we propose an information-theoretic framework for assessing the unpredictability of multi-bit strong PUF classes, demonstrating that the Entropy Rate is a pivotal metric for evaluating their resilience against ML attacks. Our proposed entropy rate estimation serves as a model-agnostic, information-theoretic lower bound on the unpredictability that holds regardless of the attack strategy used, including ML-based ones. We argue that the Uniqueness measure, defined in terms of entropy, provides a more precise and consistent evaluation compared to traditional metrics based on Hamming distance. Additionally, we present a computationally efficient method for calculating the finite-order Entropy Rate of the hybrid Boolean network (HBN) PUF, addressing the challenges posed by high dimensionality. The experimental results validate the high unpredictability and resistance of the HBN PUF class against ML attacks.