Open-RFNet: UAV RF Open-Set Recognition

Updated 8 July 2026

Open-RFNet is an open-set recognition model for UAV RF surveillance that integrates denoised preprocessing with dual-backbone feature extraction combining ResNet-18 and Transformer streams.
It employs multi-domain supervised contrastive learning to generate well-separated embeddings, significantly enhancing both closed-set accuracy and open-set rejection performance.
The system uses a GAN-based simulation with IG-OpenMax for low-overhead adapting to unknown classes, achieving high UAR and maintaining a minimal known-unknown gap.

Searching arXiv for the specified Open-RFNet paper and closely related context. Open-RFNet is an open-set recognition model for UAV radio-frequency surveillance that was introduced in "Multi-Domain Supervised Contrastive Learning for UAV Radio-Frequency Open-Set Recognition" (Gao et al., 18 Aug 2025). It is designed for the LA-ISAC setting enabled by 5G-Advanced, where unauthorized UAV activity creates a security problem for non-cooperative target monitoring. The system combines denoised RF preprocessing, dual-backbone feature extraction, multi-domain supervised contrastive learning, and an improved generative OpenMax procedure denoted IG-OpenMax. In the formulation reported by the paper, Open-RFNet targets both closed-set discrimination among known UAV types and open-set rejection of invasive UAVs that were not present during ordinary classifier training (Gao et al., 18 Aug 2025).

1. Problem setting and scope

Open-RFNet addresses UAV RF open-set recognition rather than conventional closed-set classification. In the reported experimental configuration, the dataset is DroneRFa, comprising 24 UAV types plus 1 "background"; the model uses 20 known classes and 5 held-out unknown classes, with RF collected at 915 MHz, 2.4 GHz, and 5.8 GHz by a USRP receiver (Gao et al., 18 Aug 2025). The objective is not only to classify known UAVs but also to detect samples that should be assigned to an "unknown" category at inference time.

The pipeline begins from denoised 3 ms I/Q slices that are converted by STFT into a $775\times 775$ normalized power-spectrum matrix $\hat X[n,q]$ (Gao et al., 18 Aug 2025). This preprocessing step is integral to the model definition in the paper, because denoising materially affects open-set behavior: the reported noisy-data versions attain $\mathrm{UAR}\sim 50$ – $65\%$ , whereas after denoising the reported UAR rises to $\sim 91$ – $96\%$ (Gao et al., 18 Aug 2025).

A useful way to situate Open-RFNet is as an RF-domain architecture whose core challenge is unknown-class handling. This distinguishes it from the unrelated RGB-D semantic segmentation network RFNet for road-driving images (Sun et al., 2020) and from the ReFinement Network, also abbreviated RFNet, in point cloud completion (Lyu et al., 2021). The shared acronym does not denote a shared task family.

2. System architecture

Open-RFNet consists of three major blocks: preprocessing and feature extraction, closed-set training with supervised contrastive learning, and open-set recognition via IG-OpenMax (Gao et al., 18 Aug 2025). The feature extractor is explicitly multi-stream.

The first stream extracts texture features through ResNet-18. The paper describes eight residual blocks, each following

$H_\ell = F_\ell(I_\ell;W_\ell) + I_\ell,\quad \ell=1,\dots,8,$

followed by global-average pooling and flattening to yield $\tilde z_a\in\mathbb R^{c}$ (Gao et al., 18 Aug 2025). In parallel, two TransformerEncoder-based streams extract time-frequency position information. One branch operates in the time domain and the other on $\hat X^\top$ in the frequency domain; both use an initial small MLP, sinusoidal positional encoding, stacked TransformerEncoder layers, and a terminal small MLP to produce $\tilde z_b\in\mathbb R^d$ and $\hat X[n,q]$ 0 respectively (Gao et al., 18 Aug 2025).

The three embeddings are then fused by concatenation and passed through another small MLP:

$\hat X[n,q]$ 1

The fused representation is sent to a projection head for contrastive training and to a final classification head (Gao et al., 18 Aug 2025). The paper characterizes this arrangement as a fusion of texture features and time-frequency position features from ResNet and TransformerEncoder.

Component	Function	Output
ResNet-18 branch	Texture feature extraction	$\hat X[n,q]$ 2
TransformerEncoder time branch	Time-domain position modeling	$\hat X[n,q]$ 3
TransformerEncoder frequency branch	Frequency-domain position modeling	$\hat X[n,q]$ 4

This architecture implies a deliberate separation between local texture-like RF signatures and positional structure in the time-frequency plane. A plausible implication is that the fused representation is intended to reduce failure modes associated with relying on only one RF characterization regime.

3. Multi-domain supervised contrastive learning

Closed-set training is organized around supervised contrastive shaping of the fused embedding space. After projection $\hat X[n,q]$ 5, the paper minimizes the multi-domain supervised contrastive loss

$\hat X[n,q]$ 6

where $\hat X[n,q]$ 7 is the set of anchors in a batch, $\hat X[n,q]$ 8 denotes positives with the same true label as anchor $\hat X[n,q]$ 9, $\mathrm{UAR}\sim 50$ 0 denotes the other samples, and $\mathrm{UAR}\sim 50$ 1 is a temperature hyperparameter, reported as $\mathrm{UAR}\sim 50$ 2 (Gao et al., 18 Aug 2025).

The training schedule reported in the paper is two-stage. First, the model undergoes SupCon pre-training for 30 epochs. Second, a linear softmax head is appended and the network is fine-tuned with cross-entropy for 10 more epochs (Gao et al., 18 Aug 2025). The training details list batch size $\mathrm{UAR}\sim 50$ 3 and Adam with cosine-annealing learning rate scheduling (Gao et al., 18 Aug 2025).

The paper attributes a specific representational role to the supervised contrastive objective: it pulls same-class features together and pushes different-class features apart, thereby balancing the ResNet texture stream and the Transformer-based position streams into a single embedding space with large inter-class margins (Gao et al., 18 Aug 2025). Its ablation summary is correspondingly specific. Adding the Transformer alone or SupCon alone to ResNet decreases UAR, whereas only their joint use in Open-RFNet raises UAR from approximately $\mathrm{UAR}\sim 50$ 4 to $\mathrm{UAR}\sim 50$ 5 (Gao et al., 18 Aug 2025). This suggests that the paper does not treat contrastive learning as a generic auxiliary loss, but as a mechanism that is effective only when paired with the multi-domain fusion design.

4. IG-OpenMax and open-set adaptation

The open-set component is an improved generative OpenMax algorithm, abbreviated IG-OpenMax (Gao et al., 18 Aug 2025). The reported procedure is explicitly two-stage.

In the first stage, a conditional DCGAN trained with WGAN-GP is used on the known-class feature extractor outputs to generate "unknown" RF-feature samples. The WGAN-GP objective is given as

$\mathrm{UAR}\sim 50$ 6

The paper then runs the closed-set model, denoted Open-RFNet-B, on GAN-generated samples and gathers the subset that are misclassified as simulated unknown-class features (Gao et al., 18 Aug 2025).

In the second stage, the entire feature extractor is frozen and only the final classification layer is retrained on the union of original known-class data labeled $\mathrm{UAR}\sim 50$ 7 and the simulated unknown class labeled $\mathrm{UAR}\sim 50$ 8 (Gao et al., 18 Aug 2025). The paper identifies this as the crucial low-overhead step. It further states that the rationale for freezing is that simulating unknowns at the raw-signal or full-network level is hard and unstable; by operating in the final embedding space, IG-OpenMax needs only a small adjustment in the final classification layer (Gao et al., 18 Aug 2025).

At test time, Open-RFNet computes a mean activation vector $\mathrm{UAR}\sim 50$ 9 for each class, fits a Weibull distribution $65\%$ 0 to feature-to-MAV distances, recalibrates the top- $65\%$ 1 predicted classes with $65\%$ 2, and constructs an unknown logit from the residual mass before the final softmax (Gao et al., 18 Aug 2025). The paper reports the recalibration weight for a top-ranked class $65\%$ 3 as

$65\%$ 4

with the unknown dimension handled separately (Gao et al., 18 Aug 2025).

The resulting classifier therefore combines discriminative embedding formation with EVT-style post hoc score adjustment. A plausible implication is that the method uses contrastive training to make class manifolds more separable before asking OpenMax-style recalibration to carve an explicit unknown region.

5. Data pipeline, training protocol, and complexity

The preprocessing sequence begins with I/Q slicing to 3 ms segments, corresponding to 300 k points, followed by denoising by energy threshold, STFT to a $65\%$ 5 magnitude spectrogram, and min-max normalization (Gao et al., 18 Aug 2025). Known classes are used for SupCon pre-training and closed-set fine-tuning; unknown classes appear only at test time, except for GAN-simulated samples used during the adaptation phase (Gao et al., 18 Aug 2025).

The paper’s complexity analysis reports the following values for the full Open-RFNet and a plain ResNet-18 baseline:

Model	Parameters	FLOPs per forward pass
Open-RFNet	$65\%$ 6 M ( $65\%$ 7 MB)	$65\%$ 8
Plain ResNet-18 only	$65\%$ 9 M ( $\sim 91$ 0 MB)	$\sim 91$ 1

The same section states that adding the Transformer and fusion MLPs increases parameters by approximately $\sim 91$ 2 and FLOPs by approximately $\sim 91$ 3 relative to plain ResNet-18 (Gao et al., 18 Aug 2025). This asymmetry is notable: parameter growth is dramatic, while the reported FLOP increase is moderate. The paper also notes that only the last linear classifier, now of size $\sim 91$ 4, is retrained during the open-set adaptation phase (Gao et al., 18 Aug 2025). That design choice is central to the model’s claim of low-overhead open-set adaptation.

6. Empirical performance and comparative position

The paper reports multiple performance views, and they should be distinguished carefully. In the closed-set experiment on the 20-way known-class problem, Open-RFNet achieves $\sim 91$ 5 closed-set accuracy (Gao et al., 18 Aug 2025). In the open-set experiment with 5 held-out classes, the reported values are $\sim 91$ 6, $\sim 91$ 7, and $\sim 91$ 8 (Gao et al., 18 Aug 2025). The abstract separately states that the model "achieves 95.12% in closed-set and 96.08% in open-set under 25 UAV types," so the paper presents both the 20-way closed-set accuracy figure and the KAR/UAR-based open-set view (Gao et al., 18 Aug 2025).

The open-set comparison table in the paper gives the following results:

Method	KAR	UAR
Open-RFNet (IGOM)	95.12%	96.08%
OpenMax (Ours)	95.23%	93.84%
G-OpenMax (Ours)	95.29%	93.12%
S3R-OpenSet	94.62%	88.64%
UIOS-OpenSet	93.98%	$\sim 91$ 9

The same table reports $96\%$ 0 and $96\%$ 1 for Open-RFNet, while Open-RFNet-B and Open-RFNet-G exhibit lower UAR despite comparable KAR (Gao et al., 18 Aug 2025). The paper therefore positions IG-OpenMax not as a large improvement in known-class retention, but as a stronger balancing mechanism between known and unknown recognition. That interpretation is supported by the reported gap values: $96\%$ 2 for Open-RFNet versus $96\%$ 3 for OpenMax, $96\%$ 4 for G-OpenMax, and $96\%$ 5 for S3R-OpenSet (Gao et al., 18 Aug 2025).

The t-SNE discussion in the paper further states that multi-domain SupCon disperses known classes and tightly isolates unknowns at the embedding level, enabling clean open-set boundary carving (Gao et al., 18 Aug 2025). This suggests that the quantitative gains are linked to geometry of the learned representation rather than to recalibration alone.

7. Interpretation, limitations, and nomenclatural context

Open-RFNet is best understood as a composite system in which three elements are inseparable in the reported results: multi-domain feature fusion, supervised contrastive embedding shaping, and IG-OpenMax classifier adaptation (Gao et al., 18 Aug 2025). The ablation summary argues against a common misconception that a stronger backbone or a contrastive objective alone is sufficient for RF open-set recognition; in the paper’s experiments, Transformer augmentation alone or SupCon alone decreases UAR, whereas their combination with IG-OpenMax yields the best open-set balance (Gao et al., 18 Aug 2025).

A second potential misconception is that open-set capability is learned directly from raw unknown samples. The reported method does not do this. Instead, it simulates unknowns in feature space with a conditional GAN, freezes the feature extractor, retrains only the classification head, and then applies Weibull-based recalibration at inference (Gao et al., 18 Aug 2025). The model’s unknown-class handling is therefore partly synthetic and explicitly post hoc.

Within the broader literature represented here, the name "RFNet" is ambiguous. "Real-time Fusion Network for RGB-D Semantic Segmentation Incorporating Unexpected Obstacle Detection for Road-driving Images" defines RFNet as a two-branch encoder-decoder for RGB-D road scenes (Sun et al., 2020), while "A Conditional Point Diffusion-Refinement Paradigm for 3D Point Cloud Completion" uses RFNet to mean a ReFinement Network that predicts pointwise displacements after diffusion-based coarse completion (Lyu et al., 2021). Open-RFNet is unrelated to either architecture in modality and objective: it operates on UAV RF spectrograms, uses ResNet-18 plus TransformerEncoder branches, and targets open-set recognition rather than segmentation or geometric refinement (Gao et al., 18 Aug 2025).

Taken together, the reported evidence presents Open-RFNet as a UAV RF recognition framework whose primary contribution lies in embedding-space structuring for unknown rejection. Its empirical profile is characterized by $96\%$ 6 closed-set accuracy on the 20-way known set, $96\%$ 7, $96\%$ 8, and a small reported known-unknown gap of $96\%$ 9 in the open-set setting with 25 total classes (Gao et al., 18 Aug 2025).