Temperature-Adaptive TRNG-Encrypted MRAM PUF: Enhancing Resistance to Machine Learning Attacks

Abstract

With the widespread application of the Internet-of-Things (IoT) devices, the physical unclonable function (PUF) has emerged as an essential lightweight hardware security mechanism. However, various powerful machine learning attack techniques have been developed to fake PUFs, presenting a multitude of potential risks. This article proposes a novel PUF design by using the spin-transfer torque magnetic random-access memory (STT-MRAM) and an obfuscation mechanism. The primary performance metrics in terms of uniformity (50.04%), uniqueness (49.745%), and reliability (99.59%) of the proposed PUF have been verified, which validate its functionality. Previous true random number generator (TRNG) designs based on MRAM often fail in extreme temperature conditions due to the degradation of randomness. Therefore, a temperature-adaptive TRNG (TA-TRNG) that can produce true random numbers in a broad temperature range (−25 °C to 125 ° C) is introduced in this work. These random numbers are used as “noise” to obfuscate the responses of the magnetic PUF (MPUF), which makes it difficult to crack the masked challenge-response pairs (CRPs) for attackers. Five different machine learning algorithms have been used to evaluate the resilience of the proposed TA-TRNG-PUF circuit against modeling attacks. The statistical results show that the obfuscated MPUF is on average 18.38% more immune against modeling attacks compared with the one without obfuscation, with predictions approaching random guesses (close to 50%).