International Association
for Cryptologic Research

Recent advancements in low-cost cryptography have converged upon the use of nanoscale level structural variances as sources of entropy that is unique to each device. Consequently, such delay-based Physically Unclonable Functions or (PUFs) have gained traction for several cryptographic applications. In light of recent machine learning (ML) attacks on delay-based PUFs, the common trend among PUF designers is to either introduce non-linearity using XORs or input transformations applied on the challenges in order to harden the security of delay-based PUFs. Such approaches make machine learning modelling attacks hard by destroying the linear relationship between challenge-response pairs of a PUF. However, we propose to perceive PUFs, which are fundamentally viewed as Boolean functional mapping, as a set of delay parameters drawn from normal distribution. Using this newfound perception, we propose an alternative attack strategy in this paper. We show that instead of trying to learn the exact functional relationship between challenge-response pairs from a PUF, one can search through the search space of all PUFs to find alternative PUF delay parameter set that exhibits similar behaviour as the target PUF. The core intuition behind this strategy is that one can consider a PUF as a set of stages wherein, depending on the corresponding input challenge bit, one of the several signals within a PUF's stage win a race condition. To utilize this idea, we develop a novel Particle Swarm Optimization based framework inspired by the biomimicry of amoebic reproduction. The proposed algorithm avoids the pitfalls of textbook Genetic Algorithms and demonstrates complete break of existing delay-based PUFs which are based on arbiter chains. More specifically, we are able to model higher-rder $k$-XOR PUF variants which are resistant to all-known ML modelling techniques, including $k=13, 15$ and $20$, without the knowledge of reliability values. In addition to that, we also model PUFs that incorporate input transformation, like variants of IPUF and LP-PUF. Furthermore, we take forward this idea across different search spaces in order to learn a higher order PUF using a lower order (and simpler) PUF architecture. This allows us to explore a novel class of attacks, including modelling a $k$-XOR PUF using a $(k-1)$-XOR PUF as well as bypassing input transformations based PUF designs.