The Shift PUF: Technique for Squaring the Machine Learning Complexity of Arbiter-based PUFs: Work-in-Progress

Yi Tang, Donghang Wu, Yongzhi Cao, Marian Margraf

Abstract: The physically unclonable function (PUF) is a hardware cryptographic primitive that provides identifications based on inevitable manufacturing variations, thus, as the name states, being physically unclonable. The arbiter PUF (APUF, [5] ) is a well-studied PUF design based on variational signal delays in silicon/electronic components. Using APUFs as building blocks, XOR APUF ( [12] ), lightweight secure PUF ( [9] ), feed forward APUF ( [6] , [7] ), etc. are proposed and expected to be more secure PUF designs. However, it is discovered that all of these canonical arbiter-based PUF designs suffer from machine learning modeling attacks ( [1] , [3] , [11] ). Recently, the interpose PUF (iPUF, [10] ) is proposed as a new arbiter-based PUF design that is resilient to state-of-the-art machine learning attacks. In this paper we propose a new PUF design called shift PUF that directly enhances APUF (which, to remark, is the building block of all arbiter-based PUF designs) by, as a conjecture, squaring its machine learning complexity, and consequently brings the same squaring benefit to all arbiter-based PUFs as well. To emphasize, the shift PUF itself is not a secure PUF design, and the technique of substituting APUFs with shift PUFs also not necessarily turns insecure PUF designs into secure PUF designs (the notion of security immediately follows in the next paragraph); nevertheless the technique greatly benefits already secure arbiter-based PUF designs with squared machine learning complexities.