MTJ-PUF with Input Decoder and Evaluation of Machine Learning Resistance
Takumi Kikuchi, Kimiyoshi Usami · 2024
LSIs are embedded in a wide range of products and have become an indispensable part of our lives. As a result, security has become more important than ever. Physically Unclonable Functions (PUFs), one of the LSI individual authentication techniques, have attracted attention in recent years. This is because they have potential to provide strong security by generating unique identifiers based on entropy produced during manufacturing. However, PUFs are vulnerable to machine learning (ML) modeling attacks. In this paper, we propose an approach to add an input-decoder to the Magnetic Tunnel Junction (MTJ) based latch PUF circuit to improve resistance against modeling attacks. We evaluated the vulnerability of the proposed MTJ-PUF to modeling attacks using the support vector machine (SVM), linear regression (LR), and multilayer perceptron (MLP). Results demonstrated that the prediction accuracy of each modeling attack is close to the ideal value of 50%, whereas the prediction accuracy of the conventional MTJ-PUF reaches 100% with the same number of training data. This indicates that the proposed input-decoder MTJ-PUF is resistant to ML attacks. Evaluation through simulation demonstrated that the proposed PUF achieved a uniqueness of 50%, uniformity of 49.9%, and reliability of 100%.