Zero-Shot Tag Tracking via Metadata-to-Signal Projection

Wireless devices are tracked by two independent leaks. At the digital layer, devices broadcast identifying metadata — manufacturer identifiers, service codes, transmission timings — that anyone with a radio can read. At the physical layer, tiny manufacturing imperfections in a transmitter's analog hardware imprint a distinctive distortion on the raw radio waveform, letting an eavesdropper re-identify the device even when it rotates its digital address ORACLE: Optimized Radio clAssification through Convolutional neuraL nEtworks. Every existing physical-layer tracking system, however, needs labeled waveform examples from every device it tracks: confronted with a device it has never recorded, it either fails or must first collect and cluster unlabeled signals from that device Source-Free Open-World RF Fingerprint Identification. I propose to close that gap by importing a recipe from zero-shot image classification. There, projecting class descriptions into the high-dimensional visual feature space — rather than compressing visual features down into the description space — synthesizes a prototype for a class the model has never seen, and simultaneously defuses the "hubness" failure mode in which a few points become the nearest neighbor of a disproportionate share of all queries Learning a Deep Embedding Model for Zero-Shot Learning. Applied to radio: project the openly broadcast metadata of a device into the RF feature space to synthesize its waveform prototype, and track devices whose signals have never been collected.

The proposal has two faces, one offensive and one architectural. Offensively, it quantifies a new tracking vulnerability: an eavesdropper who pairs publicly documented device metadata with waveforms from devices it has already identified can thereafter classify — and help link — never-before-recorded devices, undermining the assumption that physical-layer tracking requires prior exposure to the target. Architecturally, it brings the hubness problem itself into RF spaces. Device re-identification at scale is nearest-neighbor search in a high-dimensional embedding space over thousands of devices, exactly the regime where hubs emerge; the documented collapse of standard classifiers from 47.1% to 77.6% accuracy as populations reach 10,000 devices ChaRRNets: Channel Robust Representation Networks for RF Fingerprinting has never been examined through this lens — no reported result found on Lacuna for hubness measurements in RF embedding spaces.

Core Claim

The claim is threefold. First, that the openly observable digital side of a wireless broadcast (manufacturer and service identifiers, vendor prefixes, transmit-power fields, advertising intervals, modulation parameters) carries enough information to serve as the "semantic" side of a cross-modal mapping, and that a projection trained on seen devices to map metadata into the high-dimensional RF feature space will synthesize usable waveform prototypes for unseen devices — enabling zero-shot and generalized zero-shot device identification. Second, that nearest-prototype search in RF embedding spaces exhibits measurable hubness (a skewed distribution of how often each prototype is retrieved), that hubness grows with device population, and that the projection direction matters in radio exactly as it does in vision: mapping metadata into the RF space disperses hubs, while the reverse direction — compressing waveforms into a low-dimensional metadata space — re-creates them, echoing the 26.1% accuracy drop observed when the vision pipeline reversed its mapping direction Learning a Deep Embedding Model for Zero-Shot Learning. Third, that metadata-derived prototypes resolve a documented failure case: bit-similar hardware whose signal distortions overlap, where purely geometric methods stall Source-Free Open-World RF Fingerprint Identification.

Research Program

The first work package builds the semantic side, which is the nontrivial prerequisite: unlike image classification, wireless has no curated attribute dictionary. I will construct a structured device-description encoder from fields that are already broadcast in the clear or documented in public databases — BLE advertising payloads (manufacturer and service UUIDs, transmit power), Wi-Fi vendor prefixes and frame timing, protocol and modulation parameters, and manufacturer/model attributes derivable from FCC filings. The TPMS case shows how much identity leaks from a single static 32-bit field readable at 40 meters Security and Privacy Vulnerabilities of In-Car Wireless Networks: A Tire Pressure Monitoring System Case Study; the point here is to aggregate many such weak fields into a coherent descriptor vector. This directly targets the open problem that semantic descriptors in wireless are unstructured and noisy.

The second package trains the projection. A channel-robust encoder — the complex-valued, equivariant architecture of ChaRRNets, which models multipath channel effects as a transformation group and cancels them — produces the RF embedding, so that channel variation does not confound the cross-modal geometry. A deep model then learns $f_\theta$ mapping a device's metadata vector $s_c$ into the RF feature space, trained on seen devices with paired signals. Classification of a received signal $x$ is nearest-prototype search:

$$\hat{c} = \arg\max_c \; \cos\big(f_\theta(s_c),\, g(x)\big),$$

where $g$ is the channel-robust encoder and $f_\theta(s_c)$ is the synthesized prototype for device or device-type $c$. This mirrors prototype-based zero-shot classification in vision Convolutional Prototype Learning for Zero-Shot Recognition and the variance-preservation argument for mapping into the higher-dimensional space Crossmodal Representation Learning for Zero-shot Action Recognition; generative variants that synthesize feature distributions rather than points Generalized Zero-Shot Learning via Synthesized Examples will be compared, since synthesized distributions rather than single prototypes may better capture within-device signal variability across receivers.

The third package is a measurement study: computing the skewness of the $k$-occurrence distribution — how unevenly prototypes are retrieved as nearest neighbors — across the RF embeddings produced by existing systems on large-population benchmarks, and testing whether hubness statistically explains part of the documented scale degradation. The fourth package runs the attack and its ablations on public benchmarks, and a fifth examines defenses: how much metadata minimization (randomized payloads) degrades the attack, and whether the modality gap between metadata and signal manifolds — the very gap exploited to plant adversarial hubs in cross-modal retrieval Adversarial Hubness in Multi-Modal Retrieval — can be weaponized defensively, with a tag deliberately perturbing its waveform to become a universal hub in an eavesdropper's linkage database. That last idea is a follow-on, not the core claim.

Baselines

The do-nothing reference point is closed-set nearest-neighbor search over the seen-device RF embeddings of the same channel-robust backbone, with unseen devices simply assigned to the nearest seen prototype. On any stream containing novel devices this must collapse — unconstrained classifiers on streams where 60% of devices are novel fell below 40% H-score and AUROC in Source-Free Open-World RF Fingerprint Identification — and it establishes the floor. The key ablation isolates the projection direction: the identical backbone, metadata, and prototype search, but with the mapping reversed so that RF features are compressed into the low-dimensional metadata space where nearest-neighbor search happens. If the core claim is right, this ablation should show both worse unseen-device accuracy and elevated hubness, reproducing in radio the 26.1-point swing observed between projection directions on Animals with Attributes Learning a Deep Embedding Model for Zero-Shot Learning.

The strongest existing competitor is the source-free open-world framework of Source-Free Open-World RF Fingerprint Identification, which anchors classifier weights in an equiangular tight frame and adapts to novel devices from unlabeled signal streams; it reports an overall H-score of 97.05% on the WiSig benchmark, staying above 80% when 60% of stream devices are novel while baselines drop below 40%. It is the strongest because it is the only prior method that handles novel devices at all without source data. To be fair, it will be run with its published configuration on the same WiSig, LoRa, and Oracle splits, same backbone capacity, same signal data, and the same tuning budget as ours; our metadata is an additional input it does not use, which is precisely the ingredient under test. Beating it means matching its overall H-score while improving novel-device accuracy in the regime where no unlabeled signals from the novel devices have yet been observed — a strictly harder setting it does not address — and improving accuracy on the borderline-device subset it explicitly reports failing on.

Two further references bracket the problem rather than compete in it. The supervised CNN of ORACLE: Optimized Radio clAssification through Convolutional neuraL nEtworks reaches 99.5% on sixteen identical USRP X310 radios once intentional impairments are injected (from 35.96% without), and a median of 99% on 100 commercial Wi-Fi devices; this is the supervised ceiling that shows how much headroom remains for zero-shot methods on bit-similar hardware. On the BLE tracking side, the histogram-gradient-boosting RSSI linkage attack of RSSI-Based Fingerprinting of Bluetooth Low Energy Devices — up to 99% re-identification of randomized addresses in static settings, 1.0 accuracy with multiple receivers, but below 50% under mobility — is the shallow-learning reference for the Inria BLE experiments, and its documented mobility collapse motivates using channel-robust features.

Why not just…?

Why not just normalize the embedding space? Post-hoc corrections like Querybank Normalization or local centering are cheap fixes for hubness that a reviewer would rightly propose first. The evidence against them is that they were tested against concept-specific hubs in cross-modal retrieval and failed systematically Adversarial Hubness in Multi-Modal Retrieval, and the vision results suggest why: normalization cannot recover variance that was destroyed when features were compressed into a low-dimensional space, whereas the projection direction change from 80% to 88.1% on AwA shows the geometry can be fixed at the source Learning a Deep Embedding Model for Zero-Shot Learning. Still, this is a real threat to the geometric half of the claim, so the experiments run Querybank Normalization and local centering on the same RF embeddings with the same tuning budget; if a normalization scheme matches the projection's hubness reduction and unseen-device accuracy, the paper's claim narrows to the zero-shot capability and I will report that honestly.

Why not just inject intentional hardware impairments? ORACLE shows deliberate IQ imbalance and DC offset lift identical-radio classification from 35.96% to 99.5%, which looks like a complete solution to the borderline-device problem. It is not applicable here because it requires a cooperative software feedback loop with the transmitter — an eavesdropper studying tracking vulnerability has no such API into a stranger's tag, and cheap commercial tags lack the hardware agility. The Oracle experiment with impairments therefore serves as the supervised ceiling, not as a competitor in the threat model.

Why not just rotate identifiers at the protocol layer? MAC randomization and lightweight pseudonym schemes are the deployed defense. The record shows they do not close the leak: RSSI profiling re-identifies randomized BLE addresses at up to 99% accuracy RSSI-Based Fingerprinting of Bluetooth Low Energy Devices, and lightweight XOR/rotation token schemes for low-cost RFID are repeatedly broken by passive observers who solve the linear dependencies between sessions Vulnerability analysis of lightweight secure search protocols for low-cost RFID systems, Breaking Anonymity of Some Recent Lightweight RFID Authentication Protocols. This proposal is about the physical layer that those defenses cannot touch — and, symmetrically, its defense analysis quantifies how much metadata must also be randomized.

Why not just use a standard CNN on raw IQ samples? This is the default method and the reason the problem looks solved. It fails in exactly the regimes studied here: a standard complex-valued CNN collapses from training-day to test-day accuracy of 9.4% under channel shift (versus 44.8% for the physics-informed model), and reaches only 47.1% at 10,000 devices ChaRRNets: Channel Robust Representation Networks for RF Fingerprinting, while unconstrained classifiers in open-world streams fall below 40% H-score Source-Free Open-World RF Fingerprint Identification. The closed-set CNN will be run anyway on every split as the do-nothing reference.

Decisive Experiments

The main evaluation uses generalized-zero-shot splits of the WiSig benchmark (large-scale, cross-day, multi-receiver), the LoRa IoT dataset (25 transmitter classes), and the 10,000-device ADS-B/802.11g population from the ChaRRNet evaluation, with metrics matching the literature: H-score, AUROC, and unseen/seen accuracy. The decisive check for the core claim is unseen-device accuracy with zero signal examples of the novel classes, compared against the source-free open-world method adapted from its unlabeled stream Source-Free Open-World RF Fingerprint Identification and against the reversed-projection ablation. The borderline-device check runs on the Oracle USRP X310 setup of sixteen identical radios, where the supervised reference is 99.5% with impairments and 35.96% without ORACLE: the question is whether metadata-augmented prototypes separate bit-similar hardware that pure signal geometry cannot. The hubness check computes $k$-occurrence skewness across all embeddings at increasing population sizes, testing the correlation between hubness growth and accuracy degradation and the reduction achieved by each method. The tracking check uses the Inria BLE contact-tracing dataset, asking whether metadata prototypes extend re-identification across address randomization in the dynamic regime where RSSI classifiers fall below 50% RSSI-Based Fingerprinting of Bluetooth Low Energy Devices. Finally, a defense experiment measures the attack's degradation as metadata fields are randomized or removed, yielding a concrete standard-setting recommendation: which broadcast fields, in combination, leak enough identity to enable zero-shot physical-layer tracking.

One honest limitation deserves the last word. The projection inherits whatever the metadata does not carry: two devices with indistinguishable descriptors and indistinguishable waveforms are indistinguishable, full stop — the borderline-device experiment may well show a floor that no cross-modal method can break, and that floor is itself a finding worth publishing.