Research map / Label-efficient and robust learning / Semi-supervised learning
Open-world Radio Frequency Fingerprint Identification via Augmented Semi-supervised Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OpenRFI, an open-world RF fingerprint identification framework that combines a RoPE-augmented Informer encoder (Roinformer), SimCLR pretraining with RF-specific augmentations, and instance-level similarity and novel-class local entropy losses on top of OpenNCD.
Problem
Identifying known and unknown RF emitters in unlabeled data, where open-set methods lump all unknown emitters into one class and fail to discover and classify them individually; known-class accuracy also degrades through catastrophic forgetting.
Approach
Extends open-world new-class discovery to RF fingerprinting by using a RoPE-enhanced Informer encoder pretrained with SimCLR using AWGN and permutation augmentations, plus an instance-level InfoNCE loss that injects labeled samples and a local entropy regularizer applied to high-confidence unknown-class samples to reduce forgetting.
Key results
- OpenRFI reaches 73.95% overall accuracy on the 10-class RFF dataset vs 47.99% for OpenNCD, with novel-class accuracy 63.58%.
- On UCIHAR and SHAR, OpenRFI reaches 93.11% and 40.46% overall accuracy, versus 90.52% and 36.99% for OpenNCD.
- Roinformer with RoPE reaches 96.64% supervised test accuracy on the 32-class RFF dataset, versus 94.99% for Transformer with RoPE and 61.19% for Informer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace OpenNCD with OpenRFI | open-world recognition · RFF | overall accuracy: 47.99 -> 73.95 | Helped |
| Replace OpenNCD with OpenRFI | open-world recognition · UCIHAR | overall accuracy: 90.52 -> 93.11 | Helped |
| Replace OpenNCD with OpenRFI | open-world recognition · SHAR | overall accuracy: 36.99 -> 40.46 | Helped |
| Add RoPE positional encoding | Transformer feature extractor · 32-class RFF test set | accuracy: 52.72 -> 94.99 | Helped |
| Use Roinformer instead of Informer | feature extractor · 32-class RFF test set | accuracy: 61.19 -> 96.64 | Helped |
| Add instance-level similarity loss (ablation) | OpenRFI · 10-class RFF | overall accuracy: 21.10 -> 73.95 | Helped |
| Add local entropy regularization (ablation) | OpenRFI · 10-class RFF | seen accuracy: 57.80 -> 85.55 | Helped |
| Use modified prototype-level similarity loss L′proto (ablation) | OpenRFI · 10-class RFF | novel accuracy: 38.41 -> 63.58 | Helped |
| Set confidence threshold γ to 0.7 rather than 0.5 (ablation) | OpenRFI local entropy regularization · 10-class RFF | novel accuracy: 48.15 -> 63.58 | Helped |
| Vary confidence threshold γ (ablation) | OpenRFI local entropy regularization · 10-class RFF | overall accuracy | Mixed |
| Set local entropy regularization weight λ3 to 45 rather than 30 (ablation) | OpenRFI · 10-class RFF | novel accuracy: 56.65 -> 63.58 | Helped |
| Vary local entropy regularization weight λ3 (ablation) | OpenRFI · 10-class RFF | overall accuracy | Mixed |
Surprising findings
- Applying local entropy regularization to high-confidence known-class samples causes a cliff-like drop in known-class accuracy, and doing so for high-confidence known samples also lowers unknown-class accuracy; the authors apply it to unknown-confident samples instead.
- Plain Transformer without RoPE reaches only 52.72% accuracy, while Informer without RoPE reaches 61.19%, showing the positional encoding change matters greatly.
Limitations
- Evaluation is limited to a 10-class RFF subset in open-world setting and two human activity datasets.
- Hyperparameters (confidence threshold, entropy weight) were selected based on test-set results on the same datasets.
Experimental setup
Datasets: RFF (32-class RF fingerprint, 10-class subset for open-world) UCIHAR SHAR
Benchmarks: RFF UCIHAR SHAR
Metrics: seen accuracy novel accuracy overall accuracy test accuracy
Baselines: k-means ORCA OpenNCD Transformer Informer LegoGCD
Builds on: OpenNCD SimCLR Informer RoPE LegoGCD
Closest papers by meaning
- Towards Distance-Invariant Radio Frequency Fingerprinting via Augmented Unsupervised Learning (AAAI 2026 · similarity 0.92)
- OmniMark: Efficient and Scalable Latent Diffusion Model Fingerprinting (AAAI 2025 · similarity 0.79)
- Scalable Fingerprinting of Large Language Models (NeurIPS 2025 · similarity 0.78)
- Blind-Touch: Homomorphic Encryption-Based Distributed Neural Network Inference for Privacy-Preserving Fingerprint Authentication (AAAI 2024 · similarity 0.77)
- Physics Unrolled Neural Operator for Wireless Field Modeling (NeurIPS 2026 · similarity 0.77)
- Radio Frequency Ray Tracing with Neural Object Representation for Enhanced RF Modeling (CVPR 2025 · similarity 0.76)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.