Open Set RF Fingerprinting

• Open set RF fingerprinting is a method for physical-layer device identification that handles both known and unknown transmitters.
• It leverages deep architectures, prototype learning, and generative augmentation to distinguish and reject off-manifold signals amid hardware and channel variations.
• Evaluation protocols employ metrics such as closed-set accuracy, AUROC, and EER with EVT-based thresholding to ensure robust open-set decision-making.

Open set RF fingerprinting is the task of physical-layer device identification or authentication where, at test time, the set of transmitting devices can contain not only the identities observed during training (the “closed set”) but also previously unseen, unauthorized, or adversarial transmitters (the “open set”). Unlike closed-set classification, open-set scenarios require not only improved intra-class discrimination but robust mechanisms for confident rejection of off-manifold or out-of-distribution device fingerprints. This domain is driven by advances in deep learning, prototype/meta-learning, generative augmentation, and open-set recognition theory, and raises unique RF-specific challenges including hardware nonidealities, channel variation, and limited access to unauthorized device data.

1. Task Definition and Evaluation Protocols

The open-set RF fingerprinting problem is formally defined via two hypotheses: the sample derives from one of the $C$ known (authorized) devices, or it originates from a device outside the training set. Given a labeled dataset $$\mathcal{D} = \{(x, y) \mid x \in \mathbb{R}^T, \; y \in \{1, ..., C\}\}$$ of raw RF bursts or spectrograms, and a stream of test samples $x^*$ possibly emitted by an unknown/rogue device, the system must output either a known device identity or a robust “unknown” (“reject”) label (Wang et al., 2023, Karunaratne et al., 2021, Cai et al., 26 Jan 2025). Key metrics include closed-set accuracy, open-set (unknown) detection rate, precision-recall curves, AUROC, AUPRC, and equal error rate (EER), with thresholds commonly tuned on held-out validation sets or via extreme value modeling.

2. Architectures and Feature Learning Strategies

A range of architectures have been proposed for open-set RF fingerprinting:

• Prototype-based Learning: Methods like Improved Prototype Learning (IPL) learn a deep feature extractor $f_\theta$ and a class prototype $m_k \in \mathbb{R}^d$ per known identity, using squared Euclidean distance for logits and explicit “prototype” losses to compactify intra-class features (Wang et al., 2023).
• Auxiliary Structure: Siamese and joint-prediction architectures combine a standard classifier with a metric head, using prototypical or mean embeddings for distance-based rejection (Cai et al., 26 Jan 2025).
• Temporal Models: Architectures with CNN–LSTM stacks extract both local time–frequency features and longer-range temporal dependencies, with open-set detection driven by hidden-state distributional modeling (Puppo et al., 2023).
• Signal-processing-informed Front-Ends: Neural synchronization modules and STFT/cumulant-based preprocessing explicitly preserve hardware-unique signatures and improve network generalizability and discriminability, as seen in hypersphere-embedding discriminators (Xie et al., 2021, Wang et al., 2021, Huang et al., 26 Oct 2025).

Preprocessing steps may include IQ-to-spectrogram/STFT transforms (Wang et al., 2021, Cai et al., 26 Jan 2025), autocorrelation, channel state information (CSI) time-domain mapping (Huang et al., 26 Oct 2025), and learned synchronization (Xie et al., 2021).

3. Open-Set Decision and Calibration Techniques

Open-set recognition in RF fingerprinting leverages both classical open-set theory and recent advances:

• Distance-based Thresholding: Many systems use the minimum distance between a test embedding and all known-class prototypes/centroids, rejecting if the margin is below a calibrated threshold $\tau$ (Wang et al., 2023, Puppo et al., 2023).
• Extreme Value Theory (EVT) / OpenMax: For softmax-based classifiers, OpenMax fits Weibull distributions to tail distances from class centroids or activation vectors, recalibrates output logits, and introduces an “unknown” class. It is effective for both convolutional and time-frequency features (Wang et al., 2021, Huang et al., 26 Oct 2025). The method is robust to openness level and tail size, with best results when combining Euclidean and cosine metrics.
• Generative Augmentation of Outliers: Methods exploit variational autoencoders (VAE/CVAE) to synthesize outlier signals when real unknowns are unavailable, producing a training “negative” set to tighten open-set boundaries (Karunaratne et al., 2021).
• Dynamic Label Smoothing: Online label smoothing techniques maintain per-sample confusion histories to generate semantically soft targets and enlarge prototype separation (Wang et al., 2023).

Threshold selection and calibration strategies include per-class adaptivity (using validation means/variances of margins), EVT-based parameterization, and balancing false-reject versus false-accept rates (Wang et al., 2021, Wang et al., 2023).

4. Dataset Construction and Evaluation Methodology

Open-set RF fingerprinting is evaluated on curated datasets with well-defined authorized/unknown splits, often with real hardware diversity:

• Device Variety: Datasets range from 10–30 authorized devices (WiFi, ZigBee, LoRa) and 5–30 held-out “rogue” devices per experiment (Wang et al., 2023, Cai et al., 26 Jan 2025, Xie et al., 2021).
• Train/Test Partitioning: Known devices are split into train, validation, and test sets, with unknown devices appearing only at test time. Data augmentation (AWGN, rotations, multipath, Doppler) is widely used to increase robustness (Karunaratne et al., 2021, Cai et al., 26 Jan 2025).
• RF Signal Representation: Most experiments use fixed-length slices of either raw IQ, autocorrelation, or preamble STFTs, sometimes combined with CSI-derived time-domain vectors for commodity hardware (Huang et al., 26 Oct 2025).
• Temporal Generalization: Multi-day/multi-location captures address device drift and channel variation, with several days allocated for training and subsequent days used for “future” testing (Hanna et al., 2020, Xie et al., 2021).

Table: Example Results on Open-Set RF Fingerprinting (selected configurations)

Method	Known Acc.	Unknown Det.	Domain/Modality
IPL (Wang et al., 2023)	84.4%	94.8%	WiFi (IQ, 18 Apple)
JRFFP-SC (Cai et al., 26 Jan 2025)	98.5%	AUC=0.979	LoRa (spectrogram)
CSI²Q (Huang et al., 26 Oct 2025)	95.7%	94.6%	WiFi CSI (85 devices)
NS+BCNN+HP (Xie et al., 2021)	>99.8% AUC	<2% EER	ZigBee (IQ, multi-day)
HiNoVa (Puppo et al., 2023)	0.80–1.00	AUPRC	LoRa/WiFi (auto-LSTM)

5. Critical Design Principles and Empirical Insights

Several empirical and theoretical results have shaped current best practices:

• Feature Space Compactness and Separation: Consistency-based regularization and online label smoothing create dense, well-separated prototype clusters, reducing false acceptance of unknowns (Wang et al., 2023).
• Generative Negative Sampling: Synthetic outlier creation via VAE/CVAE or ellipsoidal/optimization-based latent sampling enables high open-set accuracy even with no real negative samples; this expands applicability to IoT scenarios with restricted hardware access (Karunaratne et al., 2021).
• Auxiliary Learning from Cross-Modality Data: Transfer learning bridges IQ-only and CSI-only paradigms, enabling OpenMax-like open-set detection using commodity WiFi hardware (Huang et al., 26 Oct 2025).
• Metric Choices: Combined Euclidean+cosine distances for activation vectors optimize class boundary modeling under open-set conditions (Wang et al., 2021).
• Complexity Management: Systems using a single prototype per class or mean-centroid reduce overfitting risk compared to many-branch per-class discriminators (Wang et al., 2023, Hanna et al., 2020).
• Temporal/Hardware Confounders: Collection over multiple sessions and environments mitigates shifts in device and channel characteristics (Hanna et al., 2020, Xie et al., 2021).

6. Open Problems, Limitations, and Prospects

Current limitations in open-set RF fingerprinting include:

• Thresholding and Calibration: Most schemes manually or per-class tune rejection thresholds based on empirical validation, which may not generalize to unseen domains, larger openness, or nonstationary SNR/channel conditions (Wang et al., 2023, Hanna et al., 2020, Huang et al., 26 Oct 2025).
• Representational Flexibility: Single-prototype models can be improved via cluster-splitting, mixture/probabilistic prototypes, or non-Euclidean geometry especially in the presence of intra-device drift or multi-modal signal distributions (Wang et al., 2023, Cai et al., 26 Jan 2025).
• Scalability and Efficiency: As the number of enrolled devices grows, maintaining high unknown detection and minimizing false rejects with limited labeled data remains challenging; some methods exhibit degraded AUC for large $\vert 𝒜\vert$ or under aggressive temporal drift (Hanna et al., 2020).
• Deployment and Adaptation: Handling multi-antenna, time-varying, or cross-band signatures is an open research direction (Xie et al., 2021, Wang et al., 2023). Integration with unsupervised adaptation and efficient hardware implementation (quantization, pruning) are underexplored.
• Metric Selection and EVT Parametrization: Optimal selection of distance metrics and EVT-tail size/tuning remains data- and architecture-dependent, motivating data-driven or meta-learned calibration strategies (Wang et al., 2021, Huang et al., 26 Oct 2025).

In summary, state-of-the-art open-set RF fingerprinting combines discriminative deep architectures with rigorous open-set modeling and, increasingly, cross-modal or generative augmentation. Prototype-based learning with regularized embeddings, dynamic calibration, and domain-adapted architectures currently enable robust unknown device rejection while sustaining high closed-set accuracy, but further improvements in threshold generalization, representational flexibility, and practical scalability remain key frontiers (Wang et al., 2023, Karunaratne et al., 2021, Huang et al., 26 Oct 2025, Cai et al., 26 Jan 2025, Xie et al., 2021, Wang et al., 2021, Puppo et al., 2023, Hanna et al., 2020).