Resilience evaluation of memristor based PUF against machine learning attacks

Physical unclonable functions (PUFs) have emerged as a favorable hardware security primitive, they exploit the process variations to provide unique signatures or secret keys amidst other critical cryptographic applications. CMOS-based PUFs are the most popular type, they generate unique bit strings using process variations in semiconductor fabrication. However, most existing CMOS PUFs are found to be vulnerable to modeling attacks based on machine learning (ML) algorithms. Memristors leveraging nanotechnology fabrication processes and highly nonlinear behavior became an interesting alternative to the existing CMOS-based PUF technology, introducing cryptographic and resilient random outputs. Memristor-based PUFs are emerging due to the inherent randomness at both the memristor level due to the cycle-to-cycle (C2C) programming variation of the device and the fabrication process level such as the cross-sectional area and variations. Our study focuses on building a machine learning analysis and attack framework of tools on <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mi>C</mml:mi> <mml:mi>u</mml:mi> <mml:mo>/</mml:mo> <mml:mi>H</mml:mi> <mml:mi>f</mml:mi> <mml:msub><mml:mi>O</mml:mi> <mml:mrow><mml:mn>2</mml:mn> <mml:mo>-</mml:mo> <mml:mi>x</mml:mi></mml:mrow> </mml:msub> <mml:mo>/</mml:mo> <mml:msup><mml:mi>p</mml:mi> <mml:mrow><mml:mo>+</mml:mo> <mml:mo>+</mml:mo></mml:mrow> </mml:msup> <mml:mi>S</mml:mi> <mml:mi>i</mml:mi></mml:mrow> </mml:math> memristor-based PUF (MR-PUF). Our objective is to test the resiliency of the security margins of the presented PUF using machine learning analysis tools, on-top of holistic NIST cryptographic randomness testing initially provided, to provide a high level of certainty in predicting the randomness output of the verified Memrister-based PUF. Our main contribution is a holistic study that focuses on attacking the randomness output resiliency based on building randomness predictors using Logistic Regression (LR), Support Vector Machine (SVM), Gaussian Mixture Models (GMM), K-means, K-means <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mo>+</mml:mo> <mml:mo>+</mml:mo></mml:mrow> </mml:math> , Random Forest, XGBoost and LSTM, within efficient time, and data complexity. Our results yield low accuracy and ROC results of within <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mn>0.49</mml:mn> <mml:mo>-</mml:mo> <mml:mn>0.52</mml:mn></mml:mrow> </mml:math> and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mn>0.49</mml:mn> <mml:mo>-</mml:mo> <mml:mn>0.52</mml:mn></mml:mrow> </mml:math> respectively, indicating failure in predicting random data demonstrates efficient randomness prediction resiliency of the MR-PUF. The efficient time and data complexities of these attacks illustrated in this study are yielded to be linear and quadratic resulting in attack execution time in seconds and 5032 training samples combined with 2157 testing samples to verify the randomness of PUF.