Physical Unclonable Functions (PUFs) exhibit asymmetric per-bit error behavior caused by manufacturing variations and aging, naturally modeled as heterogeneous Z-channels. Conventional decoders assuming symmetric BSC channels incur significant performance loss. We propose a server-side blind estimation scheme within the Reverse Fuzzy Extractor (RFE) framework, where per-bit Z-channel parameters are inferred from multi-challenge enrollment observations and incorporated into the log-likelihood ratios of a CRC-aided successive cancellation list (CA-SCL) polar decoder. We prove that single-challenge enrollment is information-theoretically insufficient for per-bit estimation. Simulations confirm that five or more enrollment challenges reduce the mean absolute estimation error sharply, achieving zero key failure rate without any device-side modification.
