A Less than 6.5E-8 BER 36-Way Reconfigurable PUF with 4-bit Stable Responses per Cell Featuring Machine Learning-Based Best-Configuration Selection
Shufan Xu, Kunyang Liu, Nan Wang, Hirofumi Shinohara, Kiichi Niitsu · 2025
Physically unclonable functions (PUFs) derive device-specific secret keys from intrinsic process variations (e.g., $V_{\text {TH }}$ mismatch). However, these small variations are vulnerable to noise and environmental fluctuations, such as supply voltage and temperature (VT) changes. Cells with small mismatches and high sensitivity to VT variations are prone to data flipping, as shown in Fig. 1(a), making it difficult to meet the strict bit error rate (BER) requirements of cryptography [1]. To address this, previous works on dark-bit masking [2]–[6] detect and discard small-mismatch cells. However, these methods often overkill otherwise stable cells, resulting in excessive bitcell loss. Reconfigurable designs [7]–[9] aim to mitigate this loss by selecting the strongest configurations from multiple candidates, but they still fail to eliminate bit errors without masking due to: (1) a limited set of configurable candidates; (2) sole reliance on mismatch while neglecting temperature sensitivity; (3) heuristic selection methods based on fixed thresholds, leading to low prediction accuracy.