PaperlayerView exact revision on arXiv

Source-linked AI summary

Physical Layer-Based Device Fingerprinting for Wireless Security: From Theory to Practice

Junqing Zhang, Francesco Ardizzon, Mattia Piana, Guanxiong Shen, Stefano Tomasin

arXiv:2506.09807v1cs.CReess.SP

TL;DR

Resource-constrained IoT devices need device-authentication approaches beyond computationally demanding cryptography. This paper surveys passive physical-layer fingerprinting through hardware impairments and channel features, covering design, applications, and research challenges. It emphasizes practical concerns including channel, noise, receiver, and hardware-pair effects, while identifying limited experimental evidence and incomplete integration as continuing challenges.

  • Problem

    Cryptographic authentication can be unsuitable for resource-constrained IoT devices because key management is challenging and public-key methods are computationally expensive.

  • Method

    The paper comprehensively surveys passive physical-layer fingerprinting through hardware impairment-based RFFI and channel-based authentication, including algorithms, experiments, and implementation considerations.

  • Results

    The survey synthesizes RFFI tasks, hardware models, deep-learning designs, channel-effect mitigation, noise and receiver-distortion mitigation, and channel-based authentication methodologies.

  • Takeaways & Limitations

    Physical-layer fingerprinting is presented as a passive authentication direction applicable to legacy IoT devices, with future work needed on integration and broader experimental validation.

Abstract

from arXiv · show

The identification of the devices from which a message is received is part of security mechanisms to ensure authentication in wireless communications. Conventional authentication approaches are cryptography-based, which, however, are usually computationally expensive and not adequate in the Internet of Things (IoT), where devices tend to be low-cost and with limited resources. This paper provides a comprehensive survey of physical layer-based device fingerprinting, which is an emerging device authentication for wireless security. In particular, this article focuses on hardware impairment-based identity authentication and channel features-based authentication. They are passive techniques that are readily applicable to legacy IoT devices. Their intrinsic hardware and channel features, algorithm design methodologies, application scenarios, and key research questions are extensively reviewed here. The remaining research challenges are discussed, and future work is suggested that can further enhance the physical layer-based device fingerprinting.

I. INTRODUCTION

IoT device authentication needs alternatives to resource-intensive cryptography, motivating passive physical-layer fingerprinting based on hardware impairments or channel characteristics. This survey reviews these approaches, their research landscape, and emerging implementation challenges.

  • Motivation: Cryptographic authentication can be difficult for IoT because key refresh is challenging and public-key algorithms impose substantial computational and power demands.The paper also notes potential vulnerability of public-key cryptography to increasing attacker computational power.
  • Physical-layer approaches: Physical-layer fingerprinting offers two passive approaches: hardware impairment-based radio-frequency fingerprint identification and channel-based authentication.RFFI uses distinctive hardware impairments, whereas channel-based authentication uses propagation characteristics associated with the transmitter’s location.
  • Research gap: Earlier surveys incompletely covered recent physical-layer device-fingerprinting technologies, channel-based methods, or in-depth machine-learning algorithms.The cited surveys variously emphasized other fingerprint types, specific application domains, or conceptual rather than detailed algorithmic coverage.
  • Research gap: Existing literature also provides limited experimental results and derivations, restricting assessment of security approaches and implementation readiness.The paper identifies missing coverage of generative AI, reconfigurable wireless environments, newer transmission bands, and cross-domain fingerprinting.
  • Survey scope: This survey comprehensively reviews RFFI and channel-based authentication, comparing their design principles, integration opportunities, practical implementation, and recent technologies.Its scope spans theoretical development, practical implementation, machine learning, reconfigurable intelligent surfaces, and multiple communication domains.

II. DEVICE FINGERPRINTING AT THE PHYSICAL LAYER

Physical-layer device fingerprinting authenticates IoT transmitters using intrinsic hardware impairments and channel features observed in received signals. Deep-learning RFFI supports several identification tasks through training and inference over extracted signal features.

  • Physical-layer fingerprinting: The receiver authenticates IoT transmitters from intrinsic hardware impairments and random channel features in received signals.Hardware effects arise from transmitter components whose specifications deviate from nominal values.
  • Signal model: The received signal combines transmitter hardware effects, multipath channel effects, receiver hardware effects, and additive white Gaussian noise.The channel impulse response may vary over time because many IoT devices are mobile.
  • Authentication features: Channel-based authentication identifies a source through propagation characteristics, whereas hardware-based RFFI exploits device-specific component imperfections.Channel features depend on propagation conditions and relative transmitter–receiver position; hardware impairments include mixer, oscillator, and amplifier effects.
  • Deep-learning RFFI tasks: Deep-learning RFFI comprises training and inference stages and can address closed-set classification, open-set recognition, and anomaly detection.These tasks differ according to whether rogue devices are involved and how they are handled.
  • Closed-set RFFI classification: In closed-set classification, a model predicts one of K legitimate device identities using a feature extractor and classifier trained on labeled packets.The final classifier layer has K outputs, and inference selects the device with the maximum probability.

B. Open-Set Recognition

Open-set recognition addresses the failure of closed-set RFFI to reject unseen rogue devices. It first distinguishes legitimate from rogue transmitters, using confidence-based or activation-vector-based methods, while related anomaly detection can reduce the task to legitimacy testing.

  • Motivation: Closed-set classification can assign an unseen rogue device to the closest legitimate class, so open-set recognition is required.The resulting admission of attackers is described as unacceptable.
  • Open-set recognition: Open-set recognition detects whether a transmitter is legitimate or rogue before classifying legitimate devices among K known identities.Rogue devices are absent from training, which gives the setting its open-set designation.
  • Confidence-based detection: Softmax-based open-set recognition flags unseen devices when prediction confidence falls below a pre-calculated threshold.Legitimate devices are handled as in closed-set classification.
  • Open-Max: Open-Max extends the activation vector to K + 1 outputs, with the additional output representing rogue devices.Unlike methods using only softmax probabilities, this approach uses the entire activation vector.
  • Anomaly detection: Anomaly detection treats legitimate devices as one class and rogue devices as another, while autoencoder methods require no other-device samples during training.The autoencoder reconstructs inputs from a compressed latent representation.
  • Transmitter impairments: Transmitter fingerprints include oscillator instability, mixer gain and phase imbalance, and power-amplifier nonlinearities.The oscillator is affected by temperature and ageing, while amplifiers exhibit AM/AM and AM/PM characteristics.

B. Receiver Impairments

Receiver impairments include local-oscillator imperfections and mixer imbalance, whose combined effects are modeled before producing the baseband signal used for RFFI. The deep-learning pipeline then preprocesses, represents, augments, and classifies sampled signals during training and inference.

  • Receiver impairment models: Receiver impairments comprise local-oscillator frequency offset and phase noise together with mixer gain and phase imbalance.These effects modify the receiver carrier and are incorporated into the received RF signal before downconversion.
  • Receiver impairment models: After downconversion and low-pass filtering, the receiver produces a baseband signal containing transmitter and receiver RF impairments.The analogue baseband signal is sampled by the ADC to produce the digital sequence used for RFFI.
  • Receiver impairment models: Using the same receiver for training and testing makes receiver-impairment effects consistent across datasets.This consistency helps prevent receiver effects from changing between the training and test conditions.
  • Deep-learning RFFI design: The deep-learning RFFI design includes dataset collection, signal preprocessing, data augmentation, signal representation, and a deep-learning model.These stages form the main processing pipeline for device identification.
  • Deep-learning RFFI design: During training, processed and represented samples are augmented before fitting the deep-learning model; during inference, the same preprocessing and representation feed the trained model.Training can usually occur offline, whereas practical inference should operate in real time.

A. Signal Collection

RFFI systems collect IQ or CSI data, preprocess and augment signals, and choose representations suited to the target model. Channel variation, unstable instantaneous CFO, and receiver or deployment constraints remain important challenges for robust identification.

  • Signal collection: RFFI datasets are collected as IQ samples using SDRs or, for some WiFi studies, CSI from commercial network interfaces and system-on-chips.Public datasets can also be used instead of building a dedicated dataset.
  • Signal preprocessing: Preprocessing commonly performs power normalization and carrier-frequency-offset compensation.CFO is the difference between transmitter and receiver carrier frequencies.
  • Signal preprocessing: Instantaneous CFO varies with temperature after power-on, although its mean remained relatively stable during a seven-month study.The stable mean was used as an auxiliary method to calibrate deep-learning predictions.
  • Data augmentation: Data augmentation generates artificial samples by adding simulated channel and noise effects, reducing the overhead of collecting comprehensive experimental datasets.The augmented signals can incorporate multipath channels, Doppler shifts, and AWGN across varied conditions.
  • Signal representation: Time-domain IQ samples are protocol-agnostic but convolve hardware features with the channel, making channel separation difficult and reducing channel-robust RFFI effectiveness.FFT coefficients and spectrograms provide alternative frequency- or time-frequency-domain representations, while other studies use DCTFs, error signals, bispectra, or Hilbert-Huang spectra.
  • Signal representation: Transient-based fingerprinting can capture power-on or power-off signatures but requires high-end receivers with high sampling rates, increasing deployment cost.Deep-learning model design should also match the signal representation, such as CNNs for image-like representations.
  • Channel effects: Wireless-channel variation can significantly degrade fingerprinting performance across WiFi, LoRa, and ZigBee signals.Multipath effects motivate channel-robust feature extraction and make channel-effect elimination a central WiFi RFFI challenge.

1) Research Activities:

Research on physical-layer device fingerprinting addresses channel, noise, and receiver effects that can obscure transmitter-specific signatures. Studies span mitigation methods, public datasets, multiple wireless technologies, and practical receiver generalization.

  • Channel effects: Channel effects, including multipath and Doppler, can significantly degrade RFFI performance and require dedicated mitigation algorithms.Signal-processing methods manually construct channel-robust features, while deep-learning methods use augmentation or learned channel-independent representations.
  • Datasets: Wideband signals are more susceptible to channel variations than narrowband signals, motivating datasets collected across diverse locations and channel conditions.Public examples include a LoRa dataset with 60 devices across six LOS and NLOS locations and a WiFi dataset with 174 transmitters and 41 receivers.
  • Noise robustness: Low SNR can submerge faint hardware fingerprints in noise, so RFFI research evaluates denoising, noise-aware training, and specialized datasets.Wireless systems often operate below 20 dB SNR; online noise augmentation has produced the most significant improvement among evaluated augmentation strategies.
  • Receiver effects: Receiver impairments can reduce identification accuracy by up to 20% when IQ imbalance differs between training and testing receivers.Proposed remedies include training with many receivers and receiver-agnostic learning; one study evaluated generalization using 20 SDR receivers.
  • Applications: Deep-learning RFFI has been applied across WiFi, ZigBee, LoRa, LTE, and satellite communications, with LoRa studies covering channel independence, low-SNR operation, receiver agnosticism, and federated protocols.For LoRa, spectrograms were identified as the most appropriate representation because of the signal’s frequency-changing property.

4) BLE:

The surveyed application landscape includes BLE, UWB, cellular, satellite, and experimental testbed platforms. Practical studies combine hardware fingerprints with varied devices, receivers, signal inputs, and collection tools.

  • BLE: BLE fingerprinting has been demonstrated for tracking mobile devices, although experiments found it viable but sometimes unreliable.A Mahalanobis-distance approach was evaluated across 17 mobile devices, including smartphones and laptops.
  • BLE: A denoising autoencoder with a CNN backbone achieved over 75% accuracy at 10 dB SNR for 18 BLE devices.The approach targets classification under low-SNR conditions.
  • UWB: UWB RFFI uses CIR measurements from commercial off-the-shelf devices because capturing its high-bandwidth signals with SDR is challenging.The cited UWB study converts CIR measurements into spectrograms for identification.
  • Cellular systems: Most recent RFFI research targets unlicensed ISM bands, while relatively few studies address cellular systems such as GSM, LTE, and 5G NR.Applications include detecting fake GSM base stations and identifying cellular transmitters from modulation and statistical features.
  • Satellite systems: Satellite RFFI has been applied to GPS spoofing detection and identification of low Earth orbit IRIDIUM satellites.Reported approaches include multivariate normal models for captured IQ samples and low-SNR satellite datasets.
  • Testbeds: RFFI testbeds use IoT kits, consumer electronics, or SDR platforms, trading controllability and hardware access against cost and practical transmission constraints.SDRs provide full hardware control but usually cost more than IoT kits and consumer electronics.
  • Signal collection: Deep-learning RFFI receivers commonly use I/Q samples captured by SDRs or CSI obtained from selected commercial WiFi chipsets and tools.CSI availability and resolution vary across tools, while I/Q processing requires synchronization and, for WiFi or Bluetooth, MAC-address decoding.

C. Requirement of Dataset Collection

Dataset collection must expose RFFI and channel-based authentication methods to realistic variation rather than relying on matched or narrowly similar conditions. The surveyed protocols differ in their assumptions about channel stability and attacker knowledge.

  • RFFI dataset requirements: RFFI evaluation can fail across different collection days even when channel conditions appear similar, showing that robustness requires temporal separation between training and testing.The cited study reports that deep-learning RFFI could not work at all under this two-day split.
  • RFFI dataset requirements: Training and testing should cover diverse channel scenarios, including LOS and NLOS, static and mobile, and indoor and outdoor conditions.Using similar environments risks overfitting when evaluating robustness to channel variation.
  • Channel-based authentication: Channel-based authentication uses measured channel features and can classify received messages as legitimate or malicious through statistical or machine-learning functions.The general scheme extracts a feature vector x and applies a classifier f(·) to determine the authentication hypothesis.
  • Authentication assumptions: Tag-based authentication assumes relatively stable channels and distinguishes transmitting locations, whereas channel-variation and challenge-response methods rely on channel changes that are predicted or controlled by the verifier.These approaches differ in whether the channel remains static, varies predictably, or is modified through a challenge.
  • Tag-based PLA: Tag-based protocols require an authenticated association phase followed by verification for each message using compatibility between the stored tag and the new channel estimate.The channel estimate is affected by mobility, noise, and receiver impairments, complicating the compatibility decision.
  • Decision methodologies: Statistical methods require knowledge of tag distributions, while machine-learning methods rely on labeled datasets; both can support binary classification when attacker information is available.Channel-variation authentication is typically framed instead as one-class classification without knowledge of attacker behavior.

B. Challenge-Response PLA

Challenge-response physical-layer authentication strengthens channel-based verification by allowing the verifier to control or vary the propagation environment. It combines channel configuration prediction with randomized challenge selection.

  • Protocol design: CR-PLA uses controllable channel configurations, such as RIS settings or drone positions, to modify the electromagnetic propagation environment.The verifier defines channel configurations according to its behavior and uses them during authentication.
  • Identification association: During association, the verifier estimates channel features across multiple configurations and interpolates features for configurations not directly explored.The initial phase must use legitimate transmissions so the learned channel model is not contaminated by spoofing.
  • Challenge and response: During verification, the verifier randomly selects a channel configuration, receives the response, and compares the measured channel with the corresponding predicted feature.An attacker must either transmit through the modified legitimate channel or know the instantaneous configuration and shape its attack accordingly.
  • Security consequence: CR-PLA introduces additional randomness into channel conditions and thereby achieves higher robustness against attacks than tag-based PLA.The surveyed literature also studies optimization of both random configuration selection and attacker strategies.
  • Channel variation: Channel-variation PLA predicts future channel features from previous observations and verifies each new packet by checking consistency with the prediction.Its predictors may use statistical methods or machine learning, with design focused on the trade-off between false alarms and missed detections.
  • Channel features: CIR and CFR provide detailed channel features for CB-PLA, but synchronization errors can introduce time, phase, or corresponding CFR phase shifts.CIR/CFR are also used in CR-PLA, including MIMO and RIS-assisted scenarios.

B. Received Power

Received power and related channel parameters provide practical features for channel-based physical-layer authentication, but their usefulness depends on spatial, temporal, and adversarial conditions. The section surveys power-, time-, angle-, and Doppler-based alternatives and their application boundaries.

  • Received-power features: Received power, also called RSS, is readily available in many receivers and is largely unaffected by small synchronization errors.Estimated SNR is analogous to received power after normalization by noise power.
  • Received-power features: Authentication can compare measured received power, RSS, or SNR with an expected reference derived from path loss and known transmitter–receiver distance.Examples include norms between measured and expected RSS, SNR, or received power, and machine-learning detection of anomalous path gains.
  • Applications: RSS supports proximity authentication, transmitter-position estimation, and channel-variation authentication across WiFi, VANET, UWAC, and VLC settings.Applications include correlated RSS for proximity, covariance-based position estimation, and line-of-sight channel gain in VLC.
  • Limitations: RSS-based localization may be vulnerable to beamforming attacks, limiting its reliability for attacker localization.The passage specifically identifies RSS as inadequate for localization under this attack condition.
  • Feature-selection boundaries: Fast channel variation can make instantaneous CIR and CFR problematic, motivating channel statistics as more robust features.Power-related features can also be too predictable to attackers or too variable because of fading.
  • Alternative channel features: Time-based features such as TOA and TDOA offer alternatives, but TDOA requires multiple synchronized receivers.AoA and Doppler provide additional position- or motion-related channel features for authentication and attacker localization.

X. IDENTIFICATION VERIFICATION METHODOLOGIES

Channel-based authentication verification is organized around statistical tests, machine-learning classifiers, and localization-oriented procedures. Statistical methods can be effective but require assumptions about feature distributions or temporal dynamics that may be difficult to obtain in practice.

  • Statistical approaches: Statistical channel-based authentication treats verification as hypothesis testing using probability density functions for legitimate and, when available, attacked conditions.The resulting test function is derived from the assumed feature distributions.
  • Statistical approaches: Likelihood-ratio testing minimizes missed detection for a fixed false-alarm probability under the Neyman–Pearson theorem.The approach requires the relevant likelihoods for the hypotheses.
  • Limitations: Statistical methods require distributional or analytic models that may be hard to obtain because feature evolution depends on both the features and the environment.Kalman filters in particular require prior knowledge of state and measurement relationships.
  • Statistical approaches: Single-sided likelihood testing uses an expected observation as a reference when only legitimate statistics are known, but is typically sub-optimal relative to likelihood-ratio testing.Examples threshold RMS error between measured and expected TDOA.
  • Statistical approaches: Pearson correlation can authenticate proximity by testing whether RSS time series from devices that should be paired are correlated.This assumes close-by transmitters produce correlated RSS sequences at a receiver.
  • Channel-variation PLA: Kalman-filter methods track evolving channel characteristics and use prediction errors, or innovations, as inputs for authenticity verification.The filter alternates prediction and model-update modes using state and measurement models.

B. ML Approaches - Binary Classification

Binary-classification approaches frame channel-based authentication as supervised learning with legitimate and attack samples, including artificially generated attacks when real attack data are unavailable. Their reported designs span conventional classifiers, neural networks, collaborative systems, and scenario-independent feature construction.

  • Problem formulation: Machine-learning authentication replaces statistical test functions with models trained on datasets containing legitimate samples, attack samples, or both.Binary classification specifically uses separate legitimate and attack datasets.
  • Theoretical result: A sufficiently complex neural network trained on sufficiently large positive and negative datasets can achieve the same performance as the optimal likelihood-ratio test.This result is stated for the tag-based physical-layer authentication formulation.
  • Classifier designs: Bagged trees achieved the best performance among decision trees, SVM, KNN, and ensemble-learning alternatives in simulations and tests, at higher computational cost.The comparison concerns four standard classification algorithms.
  • Neural architectures: Neural architectures combine feature extraction and temporal or spatial modeling, including CNN–RNN pipelines and YOLO applied to massive-MIMO angle-delay images.These designs target relevant channel features and spectral or angle-delay dependencies.
  • Limited attack data: When attack data are unavailable, binary classifiers can use artificial datasets, positive-unlabeled training, automated labeling, or synthetic samples designed from legitimate distributions.The reviewed examples include decision trees, SVMs, extreme learning, and scenario-specific negative-data construction.
  • Robust and collaborative designs: Scenario-independent authentication can use time-varying channel patterns, while other systems combine multiple receivers or devices through collaborative classification and weighted vote aggregation.These approaches address uncalibrated scenarios and distributed authentication settings.

1) Tag-based Authentication:

The tag-based authentication literature includes one-class anomaly detection, prediction, adaptive thresholding, cooperation, and attacker localization, evaluated through simulations and experimental datasets. Practical evaluation must account for hardware variation because neglecting transmitter–receiver impairments can undermine device-independent conclusions.

  • One-class classification: One-class authentication trains anomaly detectors using only legitimate data, including autoencoders, one-class SVMs, and position-confidence models.The reviewed inputs include CIR magnitude and IQ samples, while CSI can support discrete-position prediction.
  • Cooperative authentication: Cooperative one-class authentication merges data or local decisions from multiple verifiers to form a distributed authenticator.The stated goal is to improve security performance through cooperation.
  • Hybrid methods: Two-step systems first detect dataset outliers through clustering and then complete authentication with SVMs, autoencoder ensembles, or graph learning.The second stage depends on the specific reviewed implementation.
  • Comparative designs: Binary neural networks and one-class autoencoders are both used for UWAC authentication, including distributed preprocessing followed by central neural-network elaboration.The comparison appears in reviewed UWAC studies.
  • Prediction methods: Prediction-based methods use kernels, Gaussian processes, or attention autoencoders to forecast channel features, with attention accounting for fading and Doppler effects.The attention mechanism is contrasted with a traditional autoencoder predictor.
  • Adaptive and cooperative methods: Adaptive approaches update classifiers, thresholds, or protocol parameters over time using prior statistics, reinforcement learning, temporal correlation, or federated cooperation.Examples include Q-learning, Dyna-Q, deep reinforcement learning, mobile-user CSI monitoring, and edge-device collaboration.
  • Attacker localization: RSS-based clustering cannot localize multiple attackers, motivating IDOL and other methods evaluated with nearest-neighbor, probabilistic, multilateration, or neural algorithms.The limitation of the earlier clustering approach is explicitly identified before the multi-attacker framework is introduced.
  • Evaluation methodology: CB-authentication evaluation usually separates detector setup or training data from testing data, but commonly neglects hardware impairments and therefore may not establish device-independent performance.A stronger practice is to collect multiple datasets across transmitter–receiver hardware pairs or add impairment estimation and correction.

XII. CHALLENGES AND FUTURE RESEARCH

RFFI capacity, stability, and benchmarking remain insufficiently evaluated, limiting confidence in large-scale and robust device identification. The paper calls for broader experiments, theoretical analysis, and accessible datasets.

  • Lack of Capacity Evaluation: RFFI capacity—the maximum number of devices accurately distinguished by their RFFs—lacks reliable large-scale experimental and theoretical evaluation.Although one study tested up to 10,000 devices, its dataset is unpublished, and accurate capacity prediction remains difficult for deep learning-driven RFFI.
  • RFF Stability: RFF stability is insufficiently studied under environmental variation, hardware ageing, and transmitter state changes.The paper identifies robust feature extraction and evaluation beyond software-defined radios as open research directions.
  • Benchmarking: RFFI lacks a widely adopted benchmark dataset, hindering comparisons and limiting participation by researchers without RF hardware or acquisition expertise.Some public datasets exist, but most remain incomplete for broad benchmarking.

4) Limited Studies on Adversarial Machine Learning Attacks and Defense:

Physical-layer fingerprinting faces unresolved adversarial, privacy, integration, benchmarking, and generalization challenges. Future work must address attacks and defenses while extending authentication across richer scenarios and communication technologies.

  • Limited Studies on Adversarial Machine Learning Attacks and Defense: Deep-learning RFFI is exposed to training-stage backdoor attacks and inference-stage adversarial or evasion attacks.Reported attacks include perturbations that interfere with identification or induce a specific identity.
  • Limited Studies on Adversarial Machine Learning Attacks and Defense: RFFI lacks established countermeasure studies, especially for channel-propagated evasion perturbations, adversarial training, and randomized smoothing.The paper specifically notes that such countermeasures have not been studied for RFFI.
  • Privacy and Fingerprint Concealment: RFF concealment remains necessary because fingerprints intended for authentication can also enable malicious device tracking.Existing concealment approaches target phase errors or carrier-frequency offset, but deep learning may exploit additional hardware impairments.
  • Channel-Based Authentication Security: Channel-based authentication remains vulnerable to attackers transmitting from alternative positions or precoding signals to mimic legitimate channel features.The paper argues that scalable methods are needed to enlarge the current attack search space.
  • Integration: Integration of physical-layer authentication with cryptography, wiretap coding, and features from other protocol layers remains insufficiently investigated.Machine learning may help capture cross-feature relations, but resulting solutions must remain scenario-effective and robust against adversarial attacks.
  • Benchmark Datasets: CB-PLA and RFFI lack well-established benchmark datasets because evaluation spans technologies, environments, distances, positions, and attacker knowledge.More extensive data collection, spatially consistent simulators, and standardized reference scenarios could support benchmarking and adoption.
  • Generative AI Approaches: Generative AI is underused in device fingerprinting, although multimodal LLMs may generate high-fidelity artificial datasets for more robust detectors.Generative models could also be used by attackers to create samples likely to fool legitimate verifiers.
  • Emerging Communication Technologies: Device fingerprinting remains largely unexplored for newer optical communication technologies such as visible-light and underwater optical communications.The paper identifies translation of established fingerprinting approaches to these domains as a research direction.
Loading 2506.09807v1…