A machine-learning method for detecting unknown emitters aligned its learned features far more closely with simulated signal factors than the polar-input models it was compared with. But on two real-data tests, it scored lower than KAN for unknown-emitter detection.
The work is a preprint submitted to IEEE for possible publication. It was evaluated computationally on a controlled synthetic impairment benchmark and on the WiSig SingleDay and ManyTx signal subsets.
A more structured signal representation
The proposed Polar MKAN encoder uses magnitude and principal-phase inputs and connects each learned feature exclusively to one of those channels. The structure is meant to make a feature’s response easier to trace to a single part of the signal, which the paper evaluates with factor-alignment scores.
In the simulated factor-alignment experiment, Polar MKAN recorded 57.2% on disentanglement and 47.0% on completeness. Polar KAN recorded 12.9% and 13.1%, while Polar CNN scored 5.0% and 5.3%. These figures were reported as means with 95% confidence intervals across 10 runs.
That stronger alignment did not substantially reduce simulated detection. Polar MKAN reached a 73.1% ROC-AUC, a score for separating known devices from unfamiliar ones, compared with 74.4% for a raw-I/Q MKAN comparator. The paper says the difference was within one standard deviation.
The real-data trade-off
On WiSig data, Polar MKAN trailed KAN in both tested subsets. Its ROC-AUC was 78.8% versus 89.8% for KAN on SingleDay, and 67.2% versus 73.2% on ManyTx. The analysis used 28 emitters from one capture day for SingleDay and 150 emitters recorded across four days for ManyTx, using length-256 I/Q from receiver index 0.
WiSig was tested in an open-set arrangement: devices held out of training were labeled unknown at evaluation. The reported WiSig variability was calculated as standard deviation over device folds.
Response analysis supported the intended channel separation only under a wrap-free principal-phase condition: the magnitude feature stayed unchanged while both phase outputs were nondecreasing. When samples crossed the phase boundary, however, the interpretation of the phase features weakened.
A warning about phase and simulated signals
The sign of the phase features became less dependable as more simulated samples crossed that boundary. Consistency was complete when no sample wrapped, fell to roughly 70% below 2% wrapped samples, and reached about 20% to 30% once more than 10% crossed 2π. The paper reports 95% confidence intervals over 10 runs for this analysis.
Blind compensation for carrier-frequency offset, or CFO, reduced AUC for every architecture. Synthetic scores fell to about 50% to 53%, near chance, while WiSig losses were about 1 to 14 percentage points and all models remained well above chance. The authors interpret the smaller real-data drop as evidence that real emitters contain fingerprint information beyond CFO.
The evidence is limited to simulated signals and two WiSig subsets from receiver index 0. The phase guarantee is conditional on wrap-free input, and the experiments do not show that the two phase outputs isolate distinct physical impairments. The study also did not separately isolate output partitioning from monotonicity, and its fixed three-feature setting may disadvantage monotone models that require more dimensions.
Further work identified in the paper includes circular or reference-invariant phase representations, separate tests of partitioning and monotonicity, broader validation across receivers and RF datasets, and end-to-end attribution on flagged emitters. Code and configurations are reported at the project’s GitHub repository.
Paper data and sources
Original title: Interpretable Feature Learning for RF Fingerprinting via Polar MKANs
Authors: Mikhail Krasnov, Ljupcho Milosheski, Carolina Fortuna
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text