TMVS: Threshold-based Majority Voting Scheme for Robust SRAM PUFs

Using a Physically Unclonable Function (PUF) for extracting a secret key from the unique submicron structure of an integrated circuit, rather than storing it in non-volatile memory, provides notable advantages such as physical unclonability and tamper resistance. SRAM PUFs offer the advantage of reusing memories that already exist in many resource-constrained embedded devices, by leveraging the unique startup values of memory cells. However, certain SRAM cells may exhibit bit-flipping due to noise, temperature and voltage fluctuations. In this paper, we propose a straightforward yet effective method to mitigate PUF error rates. This method uses a majority voting decoder to rectify errors only in selected SRAM patterns, determined by a predefined threshold. Our scheme is simpler than other error correction codes (ECCs) based solutions. Even for the simplest ECCs, our method can provide better leftover entropy, e.g., more than 3x improvement, at the cost of worse error correction capability and memory overhead. TMVS can reach much smaller bit-error rate values compared to existing solutions based on majority voting, e.g., 360x improvement. TMVS has also minimal pre-processing runtime and cost with respect to existing pre-processing techniques like preselection and hardening.