Machine Learning Attacks on Challenge-Response Obfuscations in Strong PUFs

Physically unclonable functions (PUFs) are lightweight hardware security primitives that are used for device authentication and identification, random number generation, and cryptographic protocol design. However, it has been widely shown that PUFs can be modeled with very high accuracy using state-of-the-art machine learning (ML) and deep learning (DL) tools. Hence, recently, challenge/response obfuscation-based countermeasures have been proposed to increase the robustness of the actual PUF architectures against ML/DL attacks. In this paper, we have mainly focused on two such obfuscation-based Strong PUFs, namely Random Set-based Obfuscation PUF (RSO PUF) and ML Resilient Arbiter-PUF (Mnsl,s2,s3 PUF), published in IEEE TCAS-I'21 and VLSID‘21. In this work, we have proposed an impersonation attack on RSO PUF to extract the keys used for obfuscation, followed by modeling the design using Covariance matrix adaptation evolution strategies (CMA-ES). On the other hand, for Mnsl,s2,s3 APUF, we map its structure to work exactly like an XOR Arbiter PUF and successfully launch ML attacks such as Logistic Regression (LR), Support Vector Machines (SVM) and Multi-Layer Perceptron (MLP). Finally, we simulate both the designs by modifying the PyPUF tool and show through detailed analysis that RSO PUF with all key-set sizes (2,4,8,16, and 32) can be broken with less than 10K challenge-response pairs (CRPs) with an accuracy of 99%, whereas
$Mn_{s1,s2,s3}$
PUF can be broken with less than 150K CRPs with an accuracy of approximately 99%. In order to make sure that the attacks work in practical scenarios, we have introduced a percentage of noise (noise = 0.1 = 10%, i.e., reliability = 90%) in the simulated designs and achieved an accuracy of more than 90% for both the designs.