Large-scale deployments of resource-constrained embedded devices require lightweight, self-contained hardware roots of trust that avoid long-term secret storage. Physically Unclonable Functions (PUFs) enable secure key extraction from intrinsic hardware variations without storing keys in non-volatile memory (NVM). SRAM PUFs leverage existing Static Random Access Memory (SRAM), but face reliability issues due to environmental noise. Existing solutions rely on complex error correction codes with NVM-stored helper data, or extensive SRAM measurements to pre-select stable cells. Eliminating NVM storage for helper data mitigates information leakage risks and manufacturing costs, offering a crucial benefit for resource-constrained devices lacking NVM. This paper presents a novel approach for stabilizing SRAM PUFs without NVM-stored helper data, using a simple decoder and few SRAM measurements, at the cost of increased SRAM size. We generate consistent On-Demand Helper Data (ODHD) temporarily stored in volatile memory and validate our method experimentally on real hardware. ODHD exposes a flexible trade-off between enrollment consistency, regeneration reliability, and SRAM size by varying a single enrollment threshold, a degree of freedom absent in the fixed Dark Bit approach. At the threshold-free operating point, ODHD achieves a key error rate of 8.2% for a 16-bit output, more than six times lower than Dark Bit’s 51.5%, using only 6 bytes of SRAM per 16-bit output, when ≈ 500 power cycles are used at every key derivation.

SRAM Physically Unclonable Functions (PUFs) derive secret keys from start-up values for inherent security benefits but suffer from reliability issues due to bit flipping. We introduce the Threshold-based Majority Voting Scheme (TMVS), a lightweight method that eliminates noise and mitigates bias in SRAM PUFs while retaining the simplicity of majority voting decoders used by repetition codes, without the significant entropy loss that repetition codes incur under biased responses. TMVS runs entirely in software, requires no cell-level bit-error rate qualification or SRAM redesign, and avoids the complex decoders of heavy error correcting codes. We derive closed-form expressions for decoding-error probability and expected memory, validate them on experimental data, and present a security analysis that provides exact formulas for min-entropy and secrecy leakage due to helper data and bias, identifying conditions under which TMVS achieves zero secrecy leakage. On a large public dataset, TMVS shows near-zero cross-chip secrecy leakage and preserves average conditional min-entropy above 1 bit despite biased, spatially correlated SRAM statistics. Compared with prior work, TMVS offers the smallest decoding complexity at the cost of a larger PUF size. In a representative configuration, TMVS generates a 128-bit key with failure probability 9.15 · 10−6and zero secrecy leakage at a bit-flip probability of 10%, requiring only ∼ 248k clock cycles on a 32-bit ARM Cortex-M0. These results show that TMVS is practical and implementation-friendly for resource-constrained, low-power devices.

We study a distributed binary hypothesis testing (HT) problem with communication and security constraints, involving three parties: a remote sensor called Alice, a legitimate decision center called Bob, and an eavesdropper called Eve, all having their own source observations. In this system, Alice conveys a rate-R description of her observations to Bob, and Bob performs a binary hypothesis test on the joint distribution underlying his and Alice’s observations. The goal of Alice and Bob is to maximize the exponential decay of Bob’s miss-detection (type-II error) probability under two constraints: Bob’s false-alarm (type-I error) probability has to stay below a given threshold and Eve’s uncertainty (equivocation) about Alice’s observations should stay above a given security threshold even when Eve learns Alice’s message. For the special case of testing against independence, we characterize the largest possible type-II error exponent under the described type-I error probability and security constraints.