A sequential strong PUF architecture based on reconfigurable neural networks (RNNs) against state-of-the-art modeling attacks

Modeling attacks such as machine learning attacks are pretty efficient in breaking hardware security primitives like silicon strong physical unclonable functions (PUFs). As compared to regular strong PUFs such as arbiter PUFs and lightweight (LW)-PUFs, subthreshold current array (SCA)-PUFs exhibit a better resilience against modeling attacks since they utilize the non-linear relationship between the subthreshold current and gate voltage of transistors to obfuscate the corresponding relationship between input challenge and output response. Unfortunately, the degree of non-linearity (DoNL) within SCA-PUFs is still not sufficiently high to resist against state-of-the-art modeling attacks. Therefore, to further enhance DoNL for resisting against modeling attacks, a new strong PUF architecture is proposed in this paper by embedding reconfigurable neural networks (RNNs) into SCA-PUFs. Mathematical foundations are established for finding appropriate RNNs that are able to maximize the DoNL of an SCA-PUF. As shown in the result, when state-of-the-art modeling attacks like Lagrange multiplier attacks (LMAs) are selected, the robustness of the proposed RNN-embedded SCA-PUF is enhanced over 20 times with less than 18.5% power and area overhead as compared to a regular SCA-PUF.