Wi-Fi IoT fingerprinting gets a dual-subband gate: frozen backbones win all 72 paired runs across eight hosts with zero retraining
Synopsis
Cross-room, cross-week and cross-receiver Wi-Fi IoT radio-frequency fingerprinting has been caught in a signal-to-noise crossover: at high SNR, power-amplifier spectral regrowth on unpopulated subcarriers resists multipath but collapses as thermal noise rises, while populated preambles behave in the opposite way. GSDF-CAPC trains two separate encoders for these cues, fuses them in logit space and on a spherical embedding under a blind thermal-SNR gate calibrated only on source data, and adds label-free forward-only prototype calibration. Across DataPort-WiFi, WiSig and ORBIT MobRFFI, attaching the module to frozen backbones won all 72 paired runs over eight backbones with 0.00 percentage points of low-SNR loss.
Interpretation
Cross-domain Wi-Fi IoT radio-frequency fingerprinting can now attach the null-subcarrier spectral-regrowth cue to any frozen deployed classifier without retraining it, and fall back to the host itself at low SNR. Active-band methods previously degraded under severe indoor multipath, while methods relying only on null-subcarrier regrowth fell to near chance once thermal noise overtook intermodulation leakage, so the two cues were used separately. On the three public IEEE 802.11a/g preamble datasets DataPort-WiFi, WiSig and ORBIT MobRFFI, frozen hosts plus the module won all 72 paired runs across eight backbones on a nine-run grid, with 0.00 percentage points of low-SNR loss.
Fusion combines an independently trained null-subcarrier encoder with any frozen host at inference, with no shared parameters between the two branches. In joint training, the 20–35 dB power gap between active subcarriers and null-subcarrier regrowth lets active-band gradients dominate optimization and leaks low-SNR null-bin noise into shared hidden layers. The null-subcarrier plug-in encoder has 46.3K parameters, the same count as the native edge host; on DataPort-WiFi, removing symbol-pair coherent denoising costs about 10 percentage points under multipath while leaving the 10 dB noise result unchanged.
The fusion weight comes from a blind SNR gate calibrated only on source validation data, which estimates thermal SNR from the lag-64 periodicity of IEEE 802.11a/g preambles. Static summation injects isotropic noise below 25 dB, losing up to 44.7 percentage points at 20 dB on DataPort-WiFi, while entropy gating loses points because noise posteriors are overconfident. Averaged over eight backbones, the gate holds the same score as the host alone at 10 dB additive white Gaussian noise while retaining most of static summation's multipath gain.
Label-free target-side calibration aligns shifted prototypes with forward-only unbalanced optimal transport, degenerating to parameter-free balanced Sinkhorn when classes are roughly balanced. Backpropagation-based test-time adaptation adds gateway CPU latency and risks class collapse, while strict balanced Sinkhorn over-partitions when packet arrivals are skewed or capture windows are short. On ORBIT MobRFFI cross-receiver transfer, the balanced version reaches 52.00%, above 47.29% for five-epoch SHOT-IM and 45.33% for SHOT-Full; on skewed WiSig Day 8, the unbalanced version lifts the edge host plus fusion from 50.64% to 52.80%.
Perspective
The demonstrated scope is cross-room, cross-week and cross-receiver transfer on three public IEEE 802.11a/g preamble datasets, with single-stream preambles, closed-set devices and balanced or skewed capture windows; edge latency was measured on a single CPU thread at batch size 1 under ONNX Runtime.
Null-subcarrier regrowth fails at low SNR through quadratic decay, and the gate closes near the crossover, so gains in the transition band depend on continuous fusion rather than hard switching; on single-antenna WiSig, cross-week and cross-receiver transfer remains well below same-day cross-room results, which the authors attribute to a higher over-the-air noise floor and packet-trigger timing drift.
