What Does an RF Fingerprint Identify? A Cross-Corpus Evaluation of Hardware-Family-Associated and Unit-Level Signal
Abstract
RF fingerprinting benchmarks report how well a classifier separates transmitters, not whether that separability reflects the individual unit or its hardware family. We compare a benchmark's real hardware grouping with size-matched random groupings of the same devices, using calibrated, unclipped variational lower bounds under label-blind, device-level cross-fitting on devices never seen in training. On WiSig (OFDM Wi-Fi), the generation grouping exceeded all 999 matched random groupings (permutation p = 0.001, the resolution floor), although no grouping attained a positive certified bound, and the direction survives removing channel equalization. A preregistered replication on a LoRa corpus (chirp spread spectrum, three hardware models), with a front-end frozen on synthetic data, again exceeded all 999 (p = 0.001), and two sensitivity analyses judged by a reversal rule fixed in advance did not reverse the direction. This is a directional replication only: the excess magnitude reflects a miscalibrated null, the chain-rule decomposition leaves a residual of +0.296 bits, and capture order is undocumented. In a same-model corpus of 122 ESP32 radios, a temperature-calibrated held-out bound of +3.338 bits (label entropy 6.931 bits) indicates substantial device-specific identity, not separable from per-device capture-session effects; relying on the calibrated bound was decided after the uncalibrated bound changed sign. On synthetic transmitters a high-capacity convolutional decoder reaches 98.5% of a known 6-bit ceiling, and the audit protocol leaves chain-rule residuals of +0.072 bits (synthetic, fixed physical features) and +0.060 bits (WiSig, audit features). Two further corpora were excluded by preregistered ingest gates. Across WiSig and LoRa, hardware-family groupings consistently produced stronger identity-associated signal than size-matched random groupings.
Figures & tables
| Symbol | Definition | Role |
|---|---|---|
| coarse-label budget under a random packet split | diagnostic only; not interpretable alone | |
| coarse-label budget under a device-disjoint split | the quantity the claim rests on | |
| over size-matched random hierarchies | matched-hierarchy excess statistic |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Analysis | Cat. | Decision recorded | Result it governs |
|---|---|---|---|
| WiSig matched-hierarchy test (label-blind device folds, reference, 16D + MLP estimator) | R | protocol hash 0b4c97eaf6e1de79 ; errors in split stratification and entropy reference were corrected while WiSig was in view | final run on the frozen protocol |
| WiSig equalized = 0 sensitivity and its one-directional reading | A | e9f0eae | run after |
| SMoRFFI S1/S2/S3 and the length-metadata exclusion rule | P | 6f8547c | E11-C |
| SMoRFFI S1 implemented as a random packet-level split with the lookup-control device decoder, where the preregistration said cross-fitted bound | A | da47fdc | S1 (implementation committed before S1 was run) |
| SMoRFFI 123-device sensitivity arm | A | 2b77364 | E11-C and its sensitivity arm |
| SMoRFFI claim based on the calibrated, not the uncalibrated, bound (post-hoc estimator correction) | H | ea56d97 | decided after the uncalibrated bound read +0.514 bits on 122 devices but -0.097 bits on 123; the calibrated bound read +3.338 and +3.384 |
| Analysis | Outcome | Design fixed | Role in the paper |
|---|---|---|---|
| Synthetic ceiling and decoder saturation | proven (synthetic) | not preregistered; ground truth known by construction | shows the bound can approach a known ceiling |
| Synthetic permutation null and chain-rule residual | proven (synthetic) | not preregistered; ground truth known by construction | protocol check for the decoder |
| Synthetic cross-PHY matched-null comparison | supported (synthetic only) | before the real replication, not preregistered | motivates, does not replace, a real replication |
| Synthetic cross-PHY decomposition consistency | refuted | before the real replication, not preregistered | restricts decomposition claims to one front-end per PHY |
| WiSig generation grouping vs matched random groupings | supported (comparative) | revised with WiSig in view, then frozen | discovery |
| WiSig non-equalized preambles | supported (direction kept) | amended before the run | sensitivity |
| Number | Quantity | Estimator and role | Source |
|---|---|---|---|
| 5.910 bits (high-capacity CNN), 98.5% | held-out bound on the synthetic 6-bit ceiling | high-capacity convolutional decoder; ceiling check only | results/a17_saturation_4x.json |
| +0.072 bits | mean chain-rule residual, five seeds, synthetic | fixed 12-dimensional physical features with the same MLP decoder; protocol check, not the audit front-end | results/e1_scaling.json |
| +0.060 bits | chain-rule residual on WiSig | 16-dimensional tap-invariant features with MLP, the audit front-end; validation | results/e7_wisig_full.json |
| all 999, p = 0.001 (WiSig) | real generation grouping vs matched random groupings | frozen audit estimator; primary | results/e11_final_999.json |
| all 999, p = 0.001 (LoRa) | real model grouping vs matched random groupings | frozen dechirp front-end with the same decoder protocol; preregistered primary | results/v2_lora_e11.json |
| +0.296 bits | LoRa decomposition residual | frozen dechirp front-end; descriptive, limits interpretation | results/v2_lora_e11.json |