A machine learning resistant Arbiter PUFs scheme based on polynomial reconstruction

Physical Unclonable Functions (PUFs) are emerging as an important building block in hardware security. It has been widely used in key generation and authentication. However, Strong PUFs are vulnerable to the Machine Learning (ML) attack by modeling the challenge-response pairs (CRPs) behavior. The traditional resistant methods based on non-linear blocks (e.g. XOR) have a negative impact on the reliability of Arbiter PUFs (APUFs). In this paper, we proposed a novel error-tolerant countermeasure to resist this attack. In our Randomized Arbiter PUFs (R-APUFs) scheme, the mapping of CRPs is randomized based on the unordered feature of points in polynomial reconstruction to meet the randomness requirements of ML resistance. Moreover, the error-tolerant feature of this scheme will improve the reliability of R-APUFs when used in authentication. We exemplified the Evolution Strategy (ES)-based ML attacks on APUFs and R-APUFs. The experimental result indicates that the modeling accuracy can be reduced from over 93% to below 60%, the percentage of authentication success based on R-APUFs in 1000 tests is 100%. Therefore, this scheme can be a good supplement for the existing ML resistance work.