A Machine Learning Attack-Resilient Strong PUF Leveraging the Process Variation of MRAM

Abstract

This brief presents a parallel magnetoresistive random access memory (MRAM)-based strong physical unclonable function (MPUF). The proposed MPUF leverages fabrication-induced process variations of magnetic tunnel junction (MTJ) and compares the resistance of MRAM cells to obtain a 1-bit response value. Contrary to arbiter PUF, the proposed MPUF achieves maximum security by satisfying strict avalanche criterion (SAC) property, and its secrecy capacity increases linearly with CRPs. Moreover, an array selection circuit (ASC) is proposed, which injects strong non-linearity into the PUF response and improves the robustness against machine learning (ML)- based modeling attacks. The proposed MPUF demonstrates high reliability, uniqueness, and uniformity with mean intra and inter-hamming distance (HD) of 0.447% and 49.76%, respectively. The robustness evaluation of the proposed MPUF against ML attacks, such as multilayer perceptron (MLP), linear regression (LR), and support vector machine (SVM), shows that it is resistant to these attacks, with ML-attack prediction accuracy of less than 79.7% for the basic two-array MPUF and less than 53.8% for the four-array MPUF with ASC. Moreover, compared to existing PUFs, the proposed MPUF has a large CRP space, high energy efficiency, and low area occupancy.