Identifying Transmitters When Everyone Is Talking at Once
Radio frequency fingerprinting identifies specific transmitters by the hardware imperfections embedded in their signals. Every radio has slight manufacturing differences in its digital-to-analog converters, mixers, and power amplifiers. These imperfections create a unique signature, a fingerprint, that persists regardless of what data the radio is transmitting. Deep learning models have shown they can extract these features and classify transmitters with high accuracy, but nearly all of the existing work assumes one transmitter is active at a time.
That assumption does not hold in practice. Real wireless environments are crowded. Devices transmit on the same frequency band simultaneously, overlapping in both time and frequency. Wi-Fi routers, 4G base stations, and 5G NR radios all share the 2.4 GHz band. When multiple signals interfere, the receiver gets a superposition of distorted waveforms, and the fingerprinting model must identify which transmitters are present from this mixed signal. The problem is not just that accuracy drops under interference. It is that the entire evaluation framework breaks down when you move from single-label (one transmitter present) to multi-label (multiple transmitters present) scenarios.
Tariq Abdul-Quddoos, Xiangfang Li, and Lijun Qian at Prairie View A&M University address this by formulating co-channel RF fingerprinting as a multi-label classification problem and adding calibration guarantees on false negatives. The practical motivation is spectrum monitoring: in shared spectrum environments, regulators need to know which devices are transmitting to detect unauthorized use or policy violations. Missing a true transmitter (a false negative) is the costly error.
A Real-World Dataset From the POWDER Testbed
One of the main obstacles to studying co-channel fingerprinting is the lack of real-world datasets with multiple active emitters. Most prior work uses simulated data. Abdul-Quddoos et al. collect their dataset using the indoor over-the-air lab at the University of Utah's POWDER testbed, a 5G research platform.
The setup uses seven software-defined radios: one receiver and six transmitters. The receiver and two transmitters are NI/Ettus X310 USRPs; the remaining four are NI/Ettus B210 USRPs. Transmitted waveforms are generated by the MATLAB wireless waveform generator and consist of 5G NR, 4G LTE, and 802.11a (Wi-Fi), all with 20 MHz bandwidth at a carrier frequency of 2.425 GHz. The sampling rate is 33.33 MS/s with 20 million samples per signal.
Data is collected in three rounds, with protocol assignments rotated across radios so that each radio transmits each protocol type. For each round, four sets of data are collected at each gain level, covering every permutation of transmitter on and off states. In total, the dataset (called the POWDER Co-Channel Protocol or PCP Dataset) contains 768 signals, each representing a different combination of active transmitters and protocols. The dataset is publicly available at Hugging Face.
The signal model captures what the receiver actually sees: a superposition of active transmitter signals, each distorted by its own hardware impairments (I/Q imbalance, carrier frequency offset, carrier leakage, spurious tones, and power amplifier nonlinearity) and convolved with its own channel response, plus Gaussian noise. The distortion function for each transmitter is a cascade of I/Q imbalance, carrier-frequency and leakage effects, and power amplifier nonlinearity modeled by a memory polynomial.
Multi-Label Classification With a Lightweight CNN
The problem is formulated in two stages. First, a model takes received I/Q samples and outputs independent probabilities that each of the six known transmitters is active. Second, a post-processing function applies a threshold to these probabilities to produce a set of predicted active transmitters.
The model is a 1D convolutional neural network with 1.25 million parameters, taking up 5 MB during inference. The input is preprocessed by applying a 1024-point Discrete Fourier Transform with a Hanning window to each signal. The frequency components are split into four channels: real component, imaginary component, magnitude, and phase. The network has five 1D convolutional layers with kernel size 9, stride 1, and output channels of 256, 256, 128, 64, and 32, each followed by batch normalization, max pooling (stride and kernel size 2), and dropout (rate 0.3). The feature maps are flattened and passed through two feedforward layers, the first reducing to 256 units and the second to the number of classes (six transmitters). All layers use ReLU except the last, which uses sigmoid to produce independent per-class probabilities.
The loss function is binary cross-entropy, summing over all examples and all transmitter labels. This trains the model to output a probability for each transmitter independently, which is the correct formulation when multiple transmitters can be simultaneously active.
Calibration With Conformal Risk Control
The key contribution beyond the classifier itself is the calibration mechanism. In spectrum monitoring, the costly error is a false negative: missing a transmitter that is actually active. The model's raw probabilities do not come with guarantees about the false negative rate. Conformal risk control provides those guarantees.
The post-processing function takes the model's output probabilities and applies a threshold lambda: a transmitter is predicted active if its probability exceeds 1 minus lambda. The threshold lambda is chosen empirically from a calibration set so that the expected false negative rate is upper bounded by a user-specified value alpha. Formally, lambda-hat is the smallest value in a discrete candidate set (spaced by 0.01 from 0 to 1) such that the empirical average of the false negative rate calibration function across the calibration set is at most alpha.
The false negative rate calibration function for a single example is 1 minus the fraction of truly active transmitters that appear in the prediction set. This function is non-increasing and right-continuous, satisfying the requirements for conformal risk control. The authors calibrate for alpha values of 0.05, 0.15, and 0.25, meaning they guarantee the average false negative rate is at most 5%, 15%, or 25% respectively.
The calibration is done separately for each gain level and each number of channel occupants, reflecting the fact that channel conditions affect the model's confidence. The model is trained once on the full training set, then calibrated and evaluated at each individual operating condition.
Results: Accuracy and Calibration Under Real Interference
The evaluation uses five radios for training (X310_3, X310_4, B210_1, B210_2, B210_3) and holds out the sixth (B210_4) as out-of-distribution interference. The test set contains 10,000 examples per gain/occupancy combination, split evenly between no OOD interference and OOD interference present.
Accuracy ranges from 0.97 under favorable conditions (high gain, fewer channel occupants) down to 0.73 under challenging conditions (low gain, more occupants). The OOD interference causes a maximum accuracy loss of 0.07 and a maximum precision loss of 0.22, but the recall remains robust with a maximum loss of only 0.04.
The calibration works as intended. Micro recall scores after calibration are approximately 1 minus alpha across all conditions. At alpha = 0.05, recall is around 0.95; at alpha = 0.15, around 0.85; at alpha = 0.25, around 0.75. This holds both with and without OOD interference, demonstrating that the calibration is robust to distribution shift from an unknown interferer.
The calibration threshold lambda decreases as alpha increases (allowing more false negatives relaxes the threshold) and generally decreases as gain increases (stronger signals need less conservative thresholds). For channel occupancy, the behavior depends on alpha: at tight bounds (alpha = 0.05), lambda decreases with more occupants; at looser bounds, it increases.
Per-radio results reveal a hardware-dependent performance gap. The X310 radios, which are closer to the receiver and further from the OOD interferer, achieve precision and recall near 0.99 across all conditions. The B210 radios, which are at least 20 feet from the receiver and closer to the OOD source, show more sensitivity to gain, alpha, and interference, with precision losses up to 0.24 when OOD interference is introduced. The B210's recall is most affected by alpha, decreasing as the false negative allowance tightens.
What This Means for Shared Spectrum Monitoring
The practical value is in the calibration guarantee. A spectrum regulator monitoring a shared band can set an acceptable false negative rate and get a calibrated classifier that respects it. If the requirement is to miss no more than 5% of active transmitters on average, the conformal risk control mechanism provides that guarantee with distribution-free coverage. This is not an empirical observation about one test set; it is a statistical property that holds under the exchangeability assumption.
The robustness to OOD interference is particularly important. In a real shared spectrum environment, the interferers are not limited to the set seen during training. The fact that the calibration holds (with minor precision degradation) when an unseen radio type is introduced suggests the method generalizes beyond the specific devices used in data collection.
The lightweight model (1.25M parameters, 5 MB) is suitable for edge deployment, which matters because spectrum monitoring often needs to run on resource-constrained devices at the network edge rather than in a centralized cloud. The 1D CNN architecture avoids the computational cost of 2D processing while still capturing the frequency-domain features that carry the hardware fingerprint.
The authors note that this is a starting point. Future work will extend to calibrated anomaly detection for identifying out-of-distribution interferers with statistical guarantees, and to adjacent channel interference in addition to the co-channel scenario studied here.