MRAM-Based Physical Unclonable Function that is Resilient to Deep Learning Attacks

Electronic devices are everywhere, so strong hardware security is needed to keep private data safe from dangers. A physical Unclonable Function (PUF) device constructed using magneto-resistive random-access memory (MRAM) is shown in this study. The circuit generates appropriate outputs for problem-solving by using built-in properties derived from various manufacturing processes, such as the resistance of the magnetic tunnel junction (MTJ) cell. This proposal differs from Arbiter PUF in that it satisfies the rigorous avalanche criteria (SAC). Additionally, the proposed circuit is impervious to machine learning-based modelling assaults due to its grid-like architecture. Various ML approaches are evaluated for two- and four-array designs, including systems such as support vector machine (SVM), linear regression (LR), and multilayer perceptron (MLP). Accuracy of MLP-attack forecasts remained 53.62% for two array circuit & 49.88% for four array circuit. This demonstrates that the system continues to function well despite substantial process modifications. A convolutional neural network (CNN), a recurrent neural network (RNN), a multi-layer perceptron (MLP), and a Larq are all used in high-dimensional models that are based on deep learning. These nets have an accuracy rating of 50.32%, 50.26%, and 50.32%, along with 50.32%, correspondingly. For a simpler two-array design, the suggested circuit's efficiency at the layout stage is also looked into. Based on the simulations, the suggested circuit has an average intra-hamming distance (HD) of 0.99% and an average inter-hamming distance (HD) of 49.97%. It also has an average diffuseness of 49.08.