PaperlayerView exact revision on arXiv

Source-linked AI summary

From Identification to Discovery: Holistic RF Fingerprinting via Learning-Aided Quantization

Omer Hazan, Gil Paryanti, Nir Shlezinger

arXiv:2610.05979v1eess.SP

TL;DR

Existing RFFI methods often treat identification, unknown-device recognition, and discovery as separate tasks. This paper couples them through a learned, expandable codebook representation; experiments report accurate identification, open-set detection, and new-device discovery, including a 90% versus 40% WiSig discovery-accuracy comparison.

  • Problem

    Existing RFFI pipelines commonly separate closed-set identification, OSR, and NCD, and their representations may not expand naturally as new device classes are discovered.

    Source: §I · p. 2

  • Method

    The framework uses a learned latent representation and expandable labeled codebook to couple identification, open-set scoring, and uncertainty-guided discovery.

    Source: §I · p. 2, C. Discussion · p. 8

  • Results

    Across LoRa and WiSig, the framework supports accurate identification, open-set detection, and discovery; on WiSig with six rogue transmitters, OSRWC achieves 90% new-device accuracy versus 40% unweighted.

    Source: §V · p. 12, D. Holistic Information-Transfer… · p. 12

  • Takeaways & Limitations

    The framework transfers OSR information to discovery and expands the registered-device codebook without retraining the feature extractor as the device set evolves.

    Source: B. Training Algorithm · p. 7, C. Discussion · p. 8

Abstract

from arXiv · show

Radio frequency fingerprint identification (RFFI) enables device identification and authentication by exploiting transmitter-specific hardware impairments embedded in received RF signals. Existing RFFI methods are typically designed for a specific operating regime, such as closed-set identification, open-set recognition (OSR), or discovery of new devices, and often treat these tasks as separate algorithmic components. In this paper, we propose a holistic RFFI framework that jointly supports closed-set identification, OSR, and novel class discovery (NCD) within a single expandable architecture. The proposed method formulates RFFI as vector quantization in a learned latent feature space, where registered devices are represented by labeled codewords. This representation enables closed-set identification through codeword assignment, OSR through uncertainty measures induced by the codebook geometry, and NCD by using the latent representations and open-set uncertainty of rejected samples to organize them into new device identities. We further develop a multi-stage training procedure that combines supervised representation learning, latent-space refinement, codebook initialization, and vector-quantization-aware optimization. Experimental results on LoRa and WiFi RFFI datasets demonstrate that the proposed framework achieves accurate identification of registered devices, reliable detection of unseen transmitters, and effective discovery and registration of new device classes, highlighting the benefits of coupling the different RFFI tasks.

I. INTRODUCTION

Conventional RFFI pipelines separate closed-set identification, open-set recognition, and new-device discovery, although deployment may encounter devices absent from the design set. This paper couples these tasks through an expandable codebook-based representation and evaluates the approach on LoRa and WiFi datasets.

A. Signal Model

  • RFFI fingerprints arise from transmitter hardware impairments, enabling identification and authentication directly from received physical-layer signals rather than packet identifiers.The signal model separates the transmitter-dependent hardware transformation from channel and noise effects that may vary across conditions.Source: §I · p. 1, A. Signal Model · p. 3

I. INTRODUCTION

  • Existing pipelines separate identification, unknown-device rejection, and discovery, so adding discovered classes may require retraining or extra processing, while OSR information is rarely reused for discovery.The framework is motivated by both the difficulty of expanding a fixed-set representation and the limited transfer of information between tasks.Source: §I · p. 2
  • A holistic RFFI framework couples closed-set identification, OSR, and NCD through learned latent representations and an expandable codebook of labeled device codewords.Codebook geometry supports open-set uncertainty, which is retained to guide rejected-sample organization into new device identities.Source: §I · p. 2
  • The algorithm combines feature extraction, codebook-based identification, open-set scoring, uncertainty-aware buffering, and NCD, updating the codebook as new device classes are discovered.Its multi-stage training combines supervised learning, representation refinement, codebook initialization, and vector-quantization-aware optimization.Source: §I · p. 2
  • Experiments on LoRa and WiFi datasets report accurate registered-device identification, reliable unseen-transmitter detection, and effective discovery and incorporation of new device classes.The study evaluates all three tasks and reports benefits from the holistic design.Source: §I · p. 2

B. Problem Formulation

The formulation treats RFFI as an open-world task: systems must identify registered transmitters, recognize unseen ones, and incorporate discovered devices. It defines closed-set identification, OSR, and NCD as linked capabilities.

B. Problem Formulation

  • As deployments encounter devices outside the design-time set, a complete RFFI system must identify registered transmitters, detect unseen ones, and eventually incorporate new devices.The formulation assumes access to labeled measurements for devices in the registered set.Source: B. Problem Formulation · p. 3
  • Closed-set identification assumes every received signal comes from a registered device and estimates its identity as a multi-class classification task.Source: B. Problem Formulation · p. 3
  • OSR determines whether a received signal conforms to the registered-device distribution or should be treated as originating from an unseen transmitter.A low estimated probability of belonging to the registered set indicates an unlikely registered-device origin.Source: B. Problem Formulation · p. 3
  • NCD partitions unlabeled measurements from unseen devices into previously unseen identities, which are then registered in the known-device set.Source: B. Problem Formulation · p. 3
  • Fig. 1 illustrates the three RFFI tasks: closed-set identification, open-set recognition, and novel class discovery.Source: Figure 1 · p. 3
  • Generalized category discovery combines OSR and NCD: the system identifies registered devices, detects novel-device samples, and organizes them into new classes without prior knowledge of sample membership.The paper describes this as the open-world formulation of RFFI.Source: B. Problem Formulation · p. 3

C. Preliminaries

  • The proposed methodology builds on vector quantization and its combination with deep learning, which the paper reviews as necessary preliminaries.Source: C. Preliminaries · p. 3

1) Vector Quantization:

The preliminaries introduce vector quantization, learned quantized representations, and density-based clustering. These concepts underpin codeword assignment and clustering of device representations.

1) Vector Quantization:

  • A vector quantizer maps each d-dimensional input to a codeword in a finite codebook, typically by selecting the nearest codeword; its representation uses B = log2 |Q| bits.For d > 1, the mapping is vector quantization; Linde–Buzo–Gray learning iteratively updates codewords as cluster centroids.Source: 1) Vector Quantization · pp. 3–4

2) DL-Aided Vector Quantization:

  • A VQ-VAE maps an input through a DNN encoder to a latent representation, then divides it into M vectors that are quantized using codewords.Source: 2) DL-Aided Vector Quantization · p. 4
  • The quantized representation uses M · B bits, and a decoder maps it to an output estimate.Source: 2) DL-Aided Vector Quantization · p. 4
  • VQ-VAE training commonly combines a task-dependent reconstruction or prediction loss with codebook and commitment terms, using a stop-gradient operator and hyperparameter β.Alternative losses target improved codebook utilization or remote inference, while the cited three-term objective remains common for DNN-aided vector quantizers.Source: 2) DL-Aided Vector Quantization · p. 4

3) Density-Based Clustering:

  • DBSCAN forms clusters from density-connected core points, with samples outside sufficiently dense regions labeled as noise.A sample is a core point when its ε-neighborhood contains at least mmin samples.Source: 3) Density-Based Clustering · p. 4
  • HDBSCAN builds a hierarchy of density-based clusterings using mutual-reachability distances, then extracts persistent dense regions while retaining insufficiently supported samples as noise.Its core distance is the distance to the mminth nearest neighbor; sparse regions are effectively separated by larger mutual-reachability distances.Source: 3) Density-Based Clustering · p. 4

III. HOLISTIC RF FINGERPRINTING METHOD

The framework couples closed-set identification, open-set recognition, and novel-class discovery through an expandable labeled codebook. Its algorithm maps signals into a continuous latent space and measures their conformity to registered-device regions.

III. HOLISTIC RF FINGERPRINTING METHOD

  • Registered devices are represented by labeled codewords: nearest-codeword assignment identifies known devices, codebook geometry supports open-set recognition, and codebook expansion incorporates discovered devices.The feature extractor need not be redesigned when the codebook is augmented.Source: §III · p. 4

A. RFFI Algorithm

  • The algorithm has three stages: feature extraction, open-set score computation, and decisions for identification, rogue-device detection, or new-device discovery.The decision stage uses both the latent representation and its continuous open-set score.Source: A. RFFI Algorithm · p. 4

1) Feature Extraction:

  • The feature extractor combines handcrafted impairment features with a signal-processing branch, fusing their outputs into a continuous latent representation of device-specific characteristics.The handcrafted features include local-oscillator leakage, amplitude mismatch, and carrier frequency offset; the signal branch can use raw IQ or signal-specific preprocessing.Source: 1) Feature Extraction · pp. 4–5

2) OSR Score Computation:

  • DBSCAN models registered-device regions in latent space, and the open-set uncertainty score is the distance from a signal representation to its nearest registered-device core point.Its parameters are selected from configurations producing as many clusters as registered devices, choosing the one with greatest registered-sample coverage.Source: 2) OSR Score Computation · p. 5

3) Decision Making:

The decision stage uses the latent representation and continuous open-set score for identification, rejection, and discovery. Rejected samples retain their scores so that their degree of deviation can inform density-based clustering.

3) Decision Making:

  • The open-set threshold rejects signals whose scores exceed τ, calibrated using registered-device data without requiring examples from devices encountered only after deployment.The retained pair of latent representation and continuous score informs both the immediate decision and later discovery.Source: 3) Decision Making · p. 5
  • For accepted signals, identification assigns the device associated with the nearest labeled codeword, with multiple codewords per device accommodating intra-device variability.Source: 3) Decision Making · p. 5
  • OSRWC weights rejected samples by their open-set scores: stronger deviations receive more local replication, strengthening dense HDBSCAN regions, while borderline rejections are more likely to remain noise.This carries OSR confidence into discovery instead of treating every rejected sample equally.Source: 3) Decision Making · p. 6, Figure 3 · p. 6
  • HDBSCAN clusters the replicated rejected samples, projects labels back to originals by majority vote, and leaves originals with only noise-labeled replicas unassigned.Source: 3) Decision Making · p. 6
  • Discovered devices are incorporated by adding their codebooks and updating the registered latent set and open-set model, while keeping the feature extractor unchanged.Source: 3) Decision Making · p. 6

4) Algorithm Summary:

During operation, the algorithm identifies signals consistent with registered devices and buffers rejected signals with their continuous open-set scores. When discovery is triggered, it clusters the buffered samples and feeds discovered devices back into the registered system.

4) Algorithm Summary:

  • Each signal is mapped into latent space and scored; accepted signals use the labeled codebook, while rejected signals and their scores accumulate for discovery.The continuous scores indicate how reliable rejected samples may be for novel-class discovery.Source: 4) Algorithm Summary · p. 6
  • When a discovery episode is triggered, OSRWC supplies weighted rejected samples to HDBSCAN, whose cluster labels are projected back to the original samples.Source: 4) Algorithm Summary · pp. 6–7
  • The resulting novel classes expand the device set and codebook, closing the loop back to future identification and open-set decisions.Source: 4) Algorithm Summary · pp. 6–7

B. Training Algorithm

Training separates learning a reusable RF representation from registering the devices in a deployment. The feature extractor is pretrained and then frozen while the deployment codebook and open-set model are fitted.

B. Training Algorithm

  • Training has two phases: pretrain a device-discriminative, vector-quantization-compatible feature extractor, then freeze it and fit the codebook and open-set model to the registered-device enrollment data.The pretraining and enrollment sets may contain the same devices, but need not.Source: B. Training Algorithm · p. 7

1) Representation Pretraining:

Representation pretraining progresses from discriminative continuous features to a quantization-ready latent space. Enrollment then adapts the codebook and open-set model to registered devices while keeping the feature extractor fixed.

1) Representation Pretraining:

  • Four pretraining stages move from contrastive representation learning through classifier fitting and LBG codebook initialization to joint quantization-aware refinement.The final stage adapts the feature extractor and codebook together so quantization preserves device-discriminative structure; the feature extractor is the principal retained output.Source: 1) Representation Pretraining · p. 7
  • Supervised contrastive learning attracts representations from the same transmitter and separates those from different transmitters using augmented RF views.Augmentations include phase rotation, circular time shifting, additive noise, and amplitude scaling.Source: 1) Representation Pretraining · p. 7
  • The auxiliary classifier guides quantization-aware optimization, while LBG initializes the codebook from learned latent representations before joint refinement.The classifier and auxiliary codebook need not be retained after pretraining.Source: 1) Representation Pretraining · p. 7

2) Device Enrollment:

  • Enrollment keeps the pretrained feature extractor fixed and fits the codebook and OSR model to the registered-device set.Codewords are initialized from enrollment representations, refined with a temporary classifier, and then labeled by registered device.Source: 2) Device Enrollment · pp. 7–8

3) Training Summary:

The training and enrollment procedure pairs contrastive representation learning with quantization-aware refinement, then fits deployment-specific components to registered devices. The discussion emphasizes open-world expansion, cross-task information transfer, and computational cost.

3) Training Summary:

  • The procedure combines discriminative continuous geometry from contrastive learning with quantization-aware refinement for the discrete representation, then fits the codebook and OSR model to enrolled devices.Enrollment adapts deployment-specific components without changing the common latent representation.Source: 3) Training Summary · p. 8

C. Discussion

  • The framework couples identification, OSR, and NCD through an expandable codebook, and uses retained OSR scores to weight rejected samples during discovery.HDBSCAN needs no preset number of unseen devices and can leave insufficiently supported samples unassigned.Source: C. Discussion · p. 8
  • For 256-IQ-sample frames, increasing codebook size from P = 32 to P = 256 raises complexity only from 180.46 MFLOPs to 180.51 MFLOPs per frame.More than 99.9% of the per-frame cost comes from the convolutional feature extractor; discovery operations run only when triggered.Source: C. Discussion · p. 8
  • Possible extensions include alternative codebook-learning, quantization-aware training, density measures, and clustering that directly uses open-set confidence.Continual feature-extractor adaptation could complement codebook expansion when long-term propagation or device changes substantially alter the learned latent representation.Source: C. Discussion · p. 8

IV. EXPERIMENTAL STUDY

The study evaluates identification, open-set recognition, and discovery on LoRa and WiSig under distinct pretraining and enrollment protocols. Its HCF analysis retains three device-dependent features and excludes phase imbalance because of overlap between devices.

1) Datasets:

  • LoRa pretraining uses 30 devices and enrollment uses 10 disjoint devices; WiSig pretraining uses 50 transmitters, 10 of which are subsequently enrolled.Both protocols reserve unseen transmitters for open-set evaluation and discovery, with separate held-out samples used after registration.Source: 1) Datasets · p. 9

2) Compared Algorithms:

  • Comparisons include two RFFI-specific methods and three image-data discovery methods retrained for RF signals.Shen et al. supports closed-set identification and OSR but is not used to evaluate NCD.Source: 2) Compared Algorithms · p. 9

3) Performance Measures:

  • Closed-set performance is measured by accuracy, OSR by AUC and thresholded confusion counts, and discovery by held-out new-device accuracy.The study also measures post-discovery accuracy, clustering quality where applicable, and OSR against devices remaining unseen.Source: 3) Performance Measures · p. 9

1) HCF Evaluation:

  • CFO, amplitude imbalance, and local-oscillator leakage show transmitter-dependent behavior, whereas phase imbalance substantially overlaps between devices.The study retains the former three HCFs for subsequent experiments.Source: 1) HCF Evaluation · p. 9, Figure 4 · p. 9

2) Identification and OSR:

Across LoRa and WiSig, the framework identifies registered devices, detects unseen transmitters, and discovers new devices, with the strongest reported total accuracies after registration. Under added noise, it retains an overall post-discovery accuracy advantage in the evaluated setting.

2) Identification and OSR:

  • 97.9%–100% closed-set accuracy across the enrolled-device sweep matches or exceeds Shen et al., while ART, GCD, and OpenNCD [48] score 99.7%, 99.9%, and 99.9%.With ten enrolled devices, the method's 97.9% matches Shen et al. and exceeds OpenGCD [47] at 97.2%.Source: 2) Identification and OSR · pp. 9–10
  • 0.9947 AUC is the LoRa OSR result, ahead of 0.9531 for, 0.9436 for, 0.9417 for, 0.8929 for [47], and 0.7948 for [48].At the selected threshold, the method rejects 490 of 500 unseen samples and accepts 959 of 1000 legitimate samples; its AUC remains 0.9934–0.9967 across the rogue-device sweep.Source: 2) Identification and OSR · p. 10, Table I · p. 10, Figure 6 · p. 10

3) Generalized Category Discovery:

1) Identification and OSR:

  • In WiSig OSR, the proposed method reaches 0.9972 AUC versus 1.0000 for and 0.9968 for [47], and rejects all 175 unseen samples at the calibrated threshold.Its pre-discovery closed-set accuracy is perfect in the considered setting; only nine of 150 legitimate samples are incorrectly rejected.Source: 1) Identification and OSR · p. 10, Figure 8 · p. 10, Table III · p. 11

2) Generalized Category Discovery:

3) Robustness to Noise:

  • At 12 dB SNR, new-device accuracy is 66.2%, exceeding by 24.7 percentage points and [48] by 12.9 points, despite AUC shortfalls of 0.0167 and 0.0172.The method maintains OSR AUC above 0.97 over 30–12 dB, and Fig. 11 reports the highest mean new-device accuracy at every evaluated SNR.Source: 3) Robustness to Noise · p. 11, Figure 10 · p. 11, Figure 11 · p. 11
  • At 12 dB SNR, total post-discovery accuracy is 77.3%, versus 73.3% for [48] and 68.2% for, while previously enrolled-device accuracy is 90.3%.The discovery stage of fails in this setting.Source: 3) Robustness to Noise · p. 11

D. Holistic Information-Transfer Ablation

The ablation finds that retaining OSR information improves discovery and post-discovery identification, particularly with multiple rogue devices or degraded SNR. The conclusion presents this transfer within an expandable RFFI framework.

D. Holistic Information-Transfer Ablation

V. CONCLUSIONS

  • The conclusion describes an expandable vector-quantized latent representation that unifies closed-set identification, OSR, and NCD without requiring retraining.Newly identified devices are incorporated through codebook expansion.Source: §V · p. 12

Related papers

Loading 2610.05979v1…