Robustness Analysis of a Memristive Crossbar PUF Against Modeling Attacks

Abstract

In the greater context of computer security, hardware security issues such as integrated circuit counterfeiting, cloning, reverse engineering and piracy have emerged as critical issues due in part to an increasingly globalized supply chain. To help combat hardware security vulnerabilities, a wide range of security primitives have emerged in recent years. A popular example is physical unclonable functions (PUFs) that leverage process variations to provide unique signatures or fingerprints that can be used for authentication or secret key generation. Nanoelectronic technologies, such as the memristor technologies considered here, provide an excellent opportunity to engineer dense, energy-efficient PUF circuits with desirable statistical properties. Here, we specifically focus on the design considerations of a memristive crossbar based PUF that generates response bits as a function of variable memristor switching time. In addition to describing the operation of the crossbar PUF, we also consider its resilience to two specific machine learning attacks, specifically through the use of linear regression and support vector machines. Two circuit design modifications for the crossbar PUF are provided to improve the resilience to machine learning attacks: XORing of response bits and internal column swapping. We show that the design modifications lead to a reduction in the likelihood of successful attack to about 50% (near ideal) even given 5000 iterations for the attack itself. We also provide power estimates and performance considerations for the crossbar PUF based on three specific memristive material stacks: hafnium-oxide, tantalum-oxide, and titanium-oxide.