What Happens Before the Hash? Layer-Aware Statistical Testing in a Live QKD Pipeline
Year: 2026
Type: conference-paper
Abstract: Quantum Key Distribution (QKD) systems may expose final key material to applications through standardised delivery interfaces such as ETSI GS QKD 014, but intermediate post-processing stages—sifting, error correction, privacy amplification (PA)—are not standardised measurement points. That output-side testing can mask upstream defects is established rather than novel: AIS 31 declines to credit online tests applied to post-processed output, Owens et al. reported NIST SP 800-22 failures on raw sifted bits from an entangled-photon QKD system in 2008, and Chen et al. recently showed a Toeplitz extractor rendering an induced attack on a QRNG statistically invisible. We did not, however, find prior QKD measurements of the layer immediately upstream of the hash, which lies downstream of error correction. We instrument that boundary on a live BBM92 laboratory testbed with a minimally-invasive tap capturing the post-error-correction pay-load fed to the SHA-256 PA step and the resulting 256-bit digest, paired per record. Over a 90-minute session (N = 10,425 pairs) an implementation-checked seven-test battery at matched sample size n = 2,668,800 bits with Bonferroni correction rejects the post-EC layer on six of seven tests, the seventh (runs) failing its frequency prerequisite, while post-PA yields zero corrected rejections—a marginal-bias gap of ≈3.4 orders of magnitude. The PA input therefore remains biased after error correction. Within the session the bias shows no detectable linear trend but is over-dispersed between ≈5-minute bins, which a permutation calibration confirms is temporal rather than an artefact of record structure. We also report two specification issues in SP 800-22's longest-run test, one of which changes a verdict on our data. We scope our claims: this work does not measure the sensitivity of output-side tests to deliberately injected upstream defects, and does not certify conditional secrecy.
Institutions: Technische Universität Berlin, University of Haifa
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7223681782