A Lossless and Modeling Attack-Resistant Strong PUF with <4E-8 Bit Error Rate

Abstract

Strong physically unclonable functions (SPUFs) are promising solutions for low-cost authentication of loT edge devices, by generating an exponential number of device-specific challenge-response pairs (CRPs). Early SPUF designs are vulnerable against machine learning (ML) modeling attacks due to the lack of nonlinearity in challenge-to-response mapping [1]. Recent studies have shown that SPUFs can be designed with resiliency against ML modeling by incorporating entropy sources with non-linear operations such as Entropy LUT [2], AES S-box [3], or XOR network [4]. They achieved high resistance against known black-box ML modeling attacks with more than 0.1M training CRPs. A key challenge in these ML-resistant Strong PUF designs is ensuring the entropy sources (ES) stability under environmental variations, because a small number of unstable ES will lead to a much larger portion of unstable CRPs. The unstable CRPs need to be discarded, which reduces the number of available authentication attempts without CRP reuse. They are also a potential weak point that can be exploited to facilitate ML modeling using reliability-based attacks [5]. [2] eliminates the ES instability by hour-long accelerated aging at a high temperature, which induces a high testing cost. [3] creates an accurate ES instability map by evaluating ES under multiple temperature points and filtering out the unstable CRPs. An external access point to the ES is necessary for direct evaluation, representing another potential attack point. [4] proposes a special lithography step to randomize the interconnect, providing a more stable ES than CMOS variations. But the extra unconventional fabrication steps are undesirable in mass production.