International Association
for Cryptologic Research

Delay-based Physically Unclonable Functions (PUFs) are a popular choice for “keyless” cryptography in low-power devices. However, they have been subjected to modeling attacks using Machine Learning (ML) approaches, leading to improved PUF designs that resist ML-based attacks. On the contrary, evolutionary search (ES) based modeling approaches have garnered little attention compared to their ML counterparts due to their limited success. In this work, we revisit the problem of modeling delaybased PUFs using ES algorithms and identify drawbacks in present state-of-the-art genetic algorithms (GA) when applied to PUFs. This leads to the design of a new ES-based algorithm called CalyPSO, inspired by Particle Swarm Optimization (PSO) techniques, which is fundamentally different from classic genetic algorithm design rationale. This allows CalyPSO to avoid the pitfalls of textbook GA and mount successful modeling attacks on a variety of delay-based PUFs, including k-XOR APUF variants. Empirically, we show attacks for the parameter choices of k as high as 20, for which there are no reported ML or ES-based attacks without exploiting additional information like reliability or power/timing side-channels. We further show that CalyPSO can invade PUF designs like interpose-PUFs (i-PUFs) and (previously unattacked) LP-PUFs, which attempt to enhance ML robustness by obfuscating the input challenge. Furthermore, we evolve CalyPSO to CalyPSO++ by observing that the PUF compositions do not alter the input challenge dimensions, allowing the attacker to investigate cross-architecture modeling. This allows us to model a k-XOR APUF using a (k − 1)-XOR APUF as well as perform cross-architectural modeling of BRPUF and i-PUF using k-XOR APUF variants. CalyPSO++ provides the first modeling attack on 4 LP-PUF by reducing it to a 4-XOR APUF. Finally, we demonstrate the potency of CalyPSO and CalyPSO++ by successfully modeling various PUF architectures on noisy simulations as well as real-world hardware implementations.

We demonstrate that the Interpose PUF proposed at CHES 2019, an Arbiter PUF-based design for so-called Strong Physical Unclonable Functions (PUFs), can be modeled by novel machine learning strategies up to very substantial sizes and complexities. Our attacks require in the most difficult cases considerable, but realistic, numbers of CRPs, while consuming only moderate computation times, ranging from few seconds to few days. The attacks build on a new divide-and-conquer approach that allows us to model the two building blocks of the Interpose PUF separately. For non-reliability based Machine Learning (ML) attacks, this eventually leads to attack times on (kup, kdown)-Interpose PUFs that are comparable to the ones against max{kup, kdown}-XOR Arbiter PUFs, refuting the original claim that Interpose PUFs could provide security similar to (kdown + kup/2)-XOR Arbiter PUFs (CHES 2019). On the technical side, our novel divide-and-conquer technique might also be useful in analyzing other designs, where XOR Arbiter PUF challenge bits are unknown to the attacker.

The design of a silicon Strong Physical Unclonable Function (PUF) that is lightweight and stable, and which possesses a rigorous security argument, has been a fundamental problem in PUF research since its very beginnings in 2002. Various effective PUF modeling attacks, for example at CCS 2010 and CHES 2015, have shown that currently, no existing silicon PUF design can meet these requirements. In this paper, we introduce the novel Interpose PUF (iPUF) design, and rigorously prove its security against all known machine learning (ML) attacks, including any currently known reliability-based strategies that exploit the stability of single CRPs (we are the first to provide a detailed analysis of when the reliability based CMA-ES attack is successful and when it is not applicable). Furthermore, we provide simulations and confirm these in experiments with FPGA implementations of the iPUF, demonstrating its practicality. Our new iPUF architecture so solves the currently open problem of constructing practical, silicon Strong PUFs that are secure against state-of-the-art ML attacks.