HyDRA: A Hybrid Dual-Mode Network for Closed- and Open-Set RFFI With Optimized VMD

Hanwen Liu, Yuhe Huang, Yifeng Gong, Yanjie Zhai, Jiaxuan Lu

TL;DR

This paper introduces HyDRA, a novel hybrid network for Radio Frequency Fingerprint Identification (RFFI) that combines optimized Variational Mode Decomposition (VMD) with a dual-mode neural network (Transformer and Mamba). HyDRA achieves state-of-the-art accuracy in identifying known devices and introduces a robust open-set classification method to detect unknown devices, crucial for real-world security.

What Problem Does This Paper Address?

Device recognition is critical for wireless security, and Radio Frequency Fingerprint Identification (RFFI) offers a promising, non-cryptographic approach by leveraging unique hardware-induced signal distortions. However, existing RFFI methods face significant challenges:

  • Closed-set limitations: Traditional methods are trained and tested on a fixed set of known devices. This fails in real-world scenarios where new, unauthorized devices can appear, leading to misclassifications and security vulnerabilities.
  • Feature extraction complexity: Extracting subtle RF fingerprints requires sophisticated signal processing and deep learning techniques. Traditional methods often struggle with noise, limited scalability, and the need for extensive feature engineering.
  • Computational constraints: Deploying RFFI on edge devices requires efficient models that balance accuracy with low power consumption and fast inference times.

This paper aims to bridge these gaps by proposing a flexible and efficient RFFI system capable of both identifying known devices with high accuracy and detecting unknown devices in dynamic, real-world environments.

Methods

HyDRA is built upon several key innovations:

  • Optimized Variational Mode Decomposition (VMD): The authors propose a "lossless VMD" by fixing center frequencies and deriving closed-form solutions. This significantly reduces computational time and avoids the reconstruction errors and potential information loss associated with traditional ADMM-based VMD, leading to improved classification accuracy.
  • Hybrid Dual-Mode Network: HyDRA employs a reconfigurable neural network architecture that can dynamically switch between two powerful encoders:
  • Transformer Dynamic Sequence Encoder (TDSE): For high-precision tasks, leveraging self-attention to capture global dependencies.
  • Mamba Linear Flow Encoder (MLFE): For efficiency, using a selective state-space model (SSM) to achieve linear complexity and capture long-range dependencies, making it suitable for edge deployment.
  • Convolutional Feature Refinement Extractor (CFRE): This module preprocesses the input signal, extracting robust multi-scale temporal features by decoupling magnitude and phase processing.
  • Open-Set Classification: A novel threshold discrimination mechanism is introduced. It uses the model's softmax probability distribution and a learned threshold to distinguish between known (legal) and unknown (illegal) devices, enhancing robustness against unseen transmitters.

Datasets

The experiments were conducted on two subsets from the WiSig dataset:

  • SingleDay: A large-scale WiFi signal collection captured over a single day, containing 28 transmitters (Tx) and 10 receivers (Rx). This dataset is useful for analyzing receiver-specific variations under consistent channel conditions.
  • ManyTx: A more challenging subset spanning four capture days, including 150 Tx and 18 Rx. This dataset is used to evaluate generalization across multiple days and receivers.

Both datasets consist of IEEE 802.11a/g preamble signals, preprocessed into standardized 256-sample IQ sequences sampled at 25 Msps.

Key Insights and Ideas

  • Lossless VMD is crucial: The optimized VMD method significantly outperforms traditional VMD (ADMM) by avoiding information loss, leading to higher accuracy, especially on more complex datasets.
  • Dual-mode architecture offers flexibility: The ability to switch between TDSE (for precision) and MLFE (for efficiency) allows HyDRA to adapt to different operational needs, from high-accuracy identification to resource-constrained edge deployment.
  • Open-set classification is a necessity: The proposed threshold discrimination method effectively addresses the critical security gap of identifying unauthorized devices, a problem not adequately handled by traditional closed-set RFFI.
  • Number of IMF components matters: The study found that an optimal number of IMF components (k=3) for VMD yields the best performance, suggesting that too few or too many components can hinder feature extraction.

Key Results

  • Closed-Set Accuracy: HyDRA (TDSE) achieved state-of-the-art accuracy of 99.81% on the SingleDay dataset and 90.71% on the ManyTx dataset after lossless VMD preprocessing. HyDRA (MLFE) also demonstrated strong performance with 99.71% and 89.82% respectively.
  • Open-Set Accuracy: The proposed open-set classification method achieved a peak accuracy of 94.67% in detecting illegal devices.
  • Efficiency: HyDRA (MLFE) offers significant advantages in training speed (nearly 5.8x faster than TDSE) and model size (394 KB vs. 600 KB for TDSE), making it ideal for edge deployment.
  • Inference Speed: On the NVIDIA Jetson Xavier NX, HyDRA (TDSE) achieved an inference time of 8.4073 ms, while HyDRA (MLFE) recorded 78.2622 ms (using Mamba-minimal).

Limitations and Future Work

  • Mamba-minimal for Edge Deployment: The use of Mamba-minimal on the Jetson Xavier NX, while enabling deployment, resulted in reduced runtime efficiency compared to the official Mamba implementation.
  • Misclassification of Some Illegal Devices: The confusion matrix for open-set classification showed that some illegal devices were misclassified as legal, indicating that certain illegal devices may possess RF fingerprints similar to legal ones.
  • Future Work: The authors plan to:
  • Enable real-time signal transmission and reception via antenna integration.
  • Process and classify incoming signals online without relying on preprocessed offline datasets.
  • Develop adaptive thresholding for robust open-set classification.
  • Address challenges like real-time signal synchronization and optimizing online VMD processing for low-latency performance.

Summary updated

Abstract

Device recognition is vital for security in wireless communication systems, particularly for applications like access control. Radio Frequency Fingerprint Identification (RFFI) offers a non-cryptographic solution by exploiting hardware-induced signal distortions. This paper proposes HyDRA, a Hybrid Dual-mode RF Architecture that integrates an optimized Variational Mode Decomposition (VMD) with a novel architecture based on the fusion of Convolutional Neural Networks (CNNs), Transformers, and Mamba components, designed to support both closed-set and open-set classification tasks. The optimized VMD enhances preprocessing efficiency and classification accuracy by fixing center frequencies and using closed-form solutions. HyDRA employs the Transformer Dynamic Sequence Encoder (TDSE) for global dependency modeling and the Mamba Linear Flow Encoder (MLFE) for linear-complexity processing, adapting to varying conditions. Evaluation on public datasets demonstrates state-of-the-art (SOTA) accuracy in closed-set scenarios and robust performance in our proposed open-set classification method, effectively identifying unauthorized devices. Deployed on NVIDIA Jetson Xavier NX, HyDRA achieves millisecond-level inference speed with low power consumption, providing a practical solution for real-time wireless authentication in real-world environments.
BibTeX
@article{liu2025hydra,
  author = {Hanwen Liu and Yuhe Huang and Yifeng Gong and Yanjie Zhai and Jiaxuan Lu},
  title = {{HyDRA: A Hybrid Dual-Mode Network for Closed- and Open-Set RFFI With Optimized VMD}},
  journal = {IEEE Internet of Things Journal},
  year = {2025},
  doi = {10.1109/jiot.2025.3621243},
}
Figure