When Is the Channel a Contact Term? A Validity Test for Radio Fingerprint Tracking

Wireless devices can be tracked even after they randomize their digital identifiers. Each radio carries tiny manufacturing imperfections that leave a repeatable mark on its signal, a "radio fingerprint." The mark is mixed with the effect of the room, though: walls, distance and reflections reshape every signal before it reaches an eavesdropper. As a result, fingerprinting attacks that look nearly perfect when devices sit still degrade sharply once devices move, and nobody can say in advance which deployments are at risk. This note borrows a tool from particle physics to answer that question. When a force-carrying particle is much heavier than the energy of a collision, physicists replace it with a simple point-like "contact" interaction. They expand in the small ratio of energy to mass, and they know when that shortcut stops working. The wireless channel admits the same kind of expansion. The small ratio is the signal's bandwidth times the spread of echo delays in the room.

The core claim is this. A device's fingerprint can be split into a part that survives the channel and a part the channel destroys, and the split can be computed order by order in a single measurable number, $B\tau$. Here $B$ is the signal bandwidth and $\tau$ is the root-mean-square delay spread of the room. When $B\tau$ is small, the channel acts as a contact term that only rescales and rotates the signal. Fingerprint features that are invariant to that rescaling then transfer across locations, and the tracking risk is high. When $B\tau$ approaches one, the expansion breaks down, just as point-like dark matter models break down once the dark matter mass passes about half the mediator mass. Individual echoes must then be modeled explicitly. The result would be a predictive privacy audit, not another classifier: from a quick channel measurement, it estimates how well an eavesdropper can re-identify a device across places it was never trained on.

Background and the Gap

On the physics side, the method comes from work that integrates out heavy states. Bounds on R-parity violating supersymmetric couplings from leptonic and semileptonic meson decays reduces heavy supersymmetric exchange to static four-fermion contact terms for low-energy decays. Monochromatic gamma rays from dark matter annihilation to leptons parameterizes dark matter–lepton couplings with contact operators. It then checks them against a full model with an explicit heavy mediator and finds the contact description accurate up to roughly $m_\chi \lesssim M_{\text{med}}/2$. Mono-X versus direct searches works in the regime where that shortcut fails and the mediator must be kept as a real particle. We take three things from these papers: the expansion in a small ratio, matching against a fuller model, and an explicit validity boundary. We do not borrow any physics content.

On the tracking side, the evidence for the gap is strong but unexplained. RSSI-Based Fingerprinting of Bluetooth Low Energy Devices re-identifies Bluetooth devices with randomized addresses at up to 99% accuracy when they are static, but often below 50% when they move. Tracking UWB Devices Through Radio Frequency Fingerprinting Is Possible works with ultra-wideband channel impulse responses, meaning the received pulse together with its echoes. It reaches a 99.9% F1-score at known positions, but only 64.6% single-sample accuracy at untrained locations. A Real-World Radio Frequency Signal Dataset Based on LTE System and Variable Channels documents how non-line-of-sight channels smear hardware signatures. None of these papers predicts when a fingerprint survives movement. Each measures it after the fact in one environment.

Core Mechanism

Write the received frequency response near the carrier frequency $f_c$ as the hardware response $G(\delta)$ times the channel response $H(\delta)$, where $\delta$ is the offset from $f_c$ within the band $|\delta| \le B/2$. The channel response is a weighted sum of echoes at delays $t_i$. Its logarithm therefore has an exact expansion whose coefficients are the cumulants $\kappa_k$ of the power-weighted delay distribution:

$$\log H(\delta) = \log h_0 + \sum_{k\ge 1} \frac{(-2\pi i\,\delta)^k}{k!}\,\kappa_k .$$

This holds strictly for echoes with a common phase and approximately in general. The first cumulant is the mean delay. The second cumulant is $\tau^2$. Since $\kappa_k$ scales like $\tau^k$, the $k$-th term over the band is of size $(\pi B\tau)^k$. This is the analogue of expanding a heavy propagator in powers of momentum squared over mass squared.

Truncating at order $K$ defines a "contact channel" with $K+1$ unknowns. The order-zero term is a complex gain, the order-one term is a delay slope, and so on. Any fingerprint feature that is exactly invariant to those $K+1$ transformations carries channel contamination only of order $(\pi B\tau)^{K+1}$.

The analysis then splits two ways. When $\pi B\tau \ll 1$, low-order invariants preserve almost all of the hardware's spectral shape, so cross-location tracking should work. When $\pi B\tau \gtrsim 1$, no finite truncation suffices, and an attacker must instead estimate explicit echo taps, which is the analogue of a simplified model. The two regimes cost different amounts of the hardware signal. Dividing out more orders also removes the slowly varying parts of $G$, so in each regime there is an optimal truncation order and a predictable ceiling on accuracy.

The same lens explains the Bluetooth result. A received signal strength reading is a narrowband magnitude, so it contains only the order-zero term. That term is entirely channel times transmit power. The expansion therefore predicts that signal strength alone cannot carry a location-independent fingerprint, which matches the collapse to below 50% under mobility.

Research Program

We will first derive and implement the order-$K$ invariant features. These are the hardware spectral shape after projecting out the polynomial-in-$\delta$ components of the log response up to degree $K$. We will also derive the truncation bound as a function of $B\tau$.

Next comes matching. For each training location we fit a full multi-tap channel, the analogue of the full model with an explicit mediator. We then measure how far the order-$K$ contact channel deviates from it. This calibrates the constant in the $(\pi B\tau)^{K+1}$ law, just as the gamma-ray paper calibrated its validity bound against an explicit-mediator model.

Finally, we turn the calibrated law into a risk predictor. The predictor takes measured $B$ and $\tau$ and outputs expected cross-location re-identification accuracy and the attacker's best truncation order. We will test it on held-out locations and devices.

Baselines

The naive reference point is the same classifier trained on raw channel impulse responses with no channel handling. On the UWB data this is Scenario 2 of Tracking UWB Devices Through Radio Frequency Fingerprinting Is Possible, evaluated at untrained locations.

The strongest existing method is that paper's Vision Transformer with ArcFace loss, a metric-learning loss that pushes apart the embeddings of different devices. It reports 64.6% single-sample accuracy at untrained locations, 76.9% with ten consecutive pulses, and an open-set AUROC of 0.76. We will rerun it on identical location splits with its roughly 200,000-parameter size and the same tuning budget as our models. Beating it means reaching higher single-sample accuracy on the same untrained locations with an equal or smaller model.

The skeptic's natural competitor is full channel equalization. This divides the received spectrum by a least-squares channel estimate from the preamble, then feeds the same classifier. It is the strongest classical alternative because it removes the channel completely.

The key ablation compares three variants: fixed truncation order, the order chosen by the $B\tau$ criterion, and always using explicit echo taps. This isolates whether the validity criterion, rather than invariance itself, is what helps.

For the LTE data in A Real-World Radio Frequency Signal Dataset Based on LTE System and Variable Channels, the paper trains a CNN but we found no reported aggregate accuracy on Lacuna for XSRPdataset. We will therefore report the CNN baseline we reproduce on its 2,400/800/800 split. The Bluetooth dataset is not public, so it serves only as a qualitative check of the order-zero prediction.

Why not just…?

Why not just equalize the channel and be done? Equalization assumes nothing about scale separation. It divides out everything that varies smoothly across the band, and a per-sample channel estimate cannot tell a smooth hardware filter tilt from a channel delay slope. Our prediction is that full equalization performs well when $\pi B\tau$ is large but discards hardware signal when $\pi B\tau$ is small. In that small-$B\tau$ regime, a low-order truncation keeps the hardware's spectral curvature. The equalization baseline tests this directly. If equalization matches the selected truncation at every bandwidth, the mechanism adds nothing and we will say so.

Why not just train a bigger network with channel augmentation? This is the most serious threat. The UWB paper shows a transformer degrading to 64.6% under mobility. That number does not show augmentation would fail. We will train the same transformer with synthetic multipath augmentation drawn from the fitted tap models, under equal compute. If it matches our features, the remaining contribution is the risk predictor, which a black-box network does not supply. We would then narrow the claim to that predictor.

Why not just measure tracking accuracy in each deployment? Doing so requires an attacker-grade data collection campaign in every building. The predictor needs only a delay-spread measurement. Its value is decided by whether its predicted accuracy on held-out locations falls within a few points of measured accuracy. If it does not, the expansion is not a useful audit tool.

Proposed Experiments

The decisive experiment uses the Ruff UWB data: 14 identical boards, 50 positions and over 1.5 million pulses. We low-pass filter the impulse responses to emulate bandwidths from tens of megahertz up to the native band. This turns $B\tau$ into a controlled knob on real hardware.

The theory makes a sharp, falsifiable prediction. Plotted against $\pi B\tau$, cross-location accuracy for each truncation order should collapse onto a single curve across devices and positions. The order that maximizes accuracy should increase stepwise as $\pi B\tau$ grows. Past a crossover near $\pi B\tau \approx 1$, explicit-tap modeling should overtake every finite truncation, mirroring the $m_\chi \approx M_{\text{med}}/2$ boundary.

On the LTE dataset, line-of-sight and non-line-of-sight captures supply two natural $B\tau$ regimes. The theory predicts that the gap between them shrinks under the selected truncation and widens under raw features.

Three outcomes would falsify the idea. The first is curves that do not collapse with $B\tau$. The second is a predictor error larger than the spread between baselines. The third is augmentation matching our features while also producing calibrated risk estimates.