TL;DR
This paper introduces TWC-CNN, a novel method for Specific Emitter Identification (SEI) that excels in low-data scenarios. By combining continuous wavelet transform (CWT) for data augmentation with a dual-modality neural network and contrastive learning, TWC-CNN achieves high accuracy even with as little as 1% of training data. This is crucial for identifying short-duration or non-cooperative transmitters where data is scarce.
What Problem Does This Paper Address?
Identifying the unique physical layer characteristics of radio transmitters, known as Specific Emitter Identification (SEI), is vital for security applications like authentication and detecting illegal transmissions. While deep learning has significantly boosted SEI performance, a major hurdle remains: many real-world scenarios, especially those involving non-cooperative or transient transmitters, provide very limited training data. Training deep learning models on such sparse datasets often leads to underfitting and poor accuracy. Traditional SEI methods also struggle with this "small sample" problem. This paper aims to bridge this gap by developing an SEI method that is robust and accurate even when faced with a severe scarcity of training samples.
Methods
The core of the proposed TWC-CNN method lies in its innovative approach to data augmentation and feature extraction:
- Time-Wavelet Spectrum Consistency (TWC) Data Augmentation: Instead of traditional methods like adding noise or simple transformations, the authors employ Continuous Wavelet Transform (CWT) on the in-phase and quadrature (IQ) data. This transforms the 1D signal into a 2D time-wavelet spectrum, creating a new "modality" of data. This process is applied to the original samples, generating augmented samples that are significantly different from the originals, thus expanding the effective training dataset.
- Dual-Modality Feature Extraction: The method uses two separate neural network branches to process the original IQ data and the CWT-generated wavelet spectra. A Complex-Valued Neural Network (CVNN) handles the IQ data, preserving the complex nature of radio signals, while a Deep Convolutional Neural Network (DCNN) processes the 2D wavelet spectra. Both networks extract 1024-dimensional feature vectors.
- Contrastive Learning Framework: To ensure consistency between the features extracted from the two modalities (IQ and wavelet spectrum) and to improve generalization with limited data, the authors employ a combination of losses:
- NT-Xent Loss: Aims to pull features from the same emitter (original and augmented samples) closer together in the feature space.
- Cross-Entropy Loss (CE): Standard supervised loss for classification.
- Cosine Loss: Used with a "vector-mask module" to enforce similarity between the feature distributions of the original and augmented samples, correcting potential discrepancies introduced by augmentation.
- Pre-training and Fine-tuning: The network is first pre-trained using the combined loss functions to learn robust feature representations. Subsequently, a fine-tuning stage uses a fully connected classifier on concatenated features for the final SEI task.
Datasets
The proposed TWC-CNN method was evaluated on two publicly available datasets:
- ADS-B (Automatic Dependent Surveillance–Broadcast) Dataset: Contains signals from civil and military aircraft, providing information like position, altitude, and speed.
- WiFi Dataset: Generated using USRP transmitters and compliant with the IEEE 802.11a standard.
The paper specifically focuses on simulating "small sample" SEI scenarios by creating datasets with varying percentages of training data (1%, 5%, 10%, 15%, 20%) from the original datasets.
Key Insights and Ideas
- CWT as a Powerful Data Augmentation Tool: The paper demonstrates that CWT is not just a simple augmentation but a transformative process that creates a new, complementary data modality (time-wavelet spectrum). This richer representation aids in feature extraction.
- Multimodal Learning for SEI: By processing both the original IQ data and the CWT-generated spectra with separate networks and then combining their learned features, the method effectively leverages information from different perspectives of the radio signal.
- Contrastive Learning's Role in Low-Data SEI: The combination of NT-Xent and cosine loss is crucial for aligning features from original and augmented data, enabling the model to learn discriminative representations even with very few labeled samples. This helps overcome the underfitting problem.
- The Importance of Feature Distribution Consistency: The cosine loss component specifically addresses the potential mismatch in feature distributions between original and augmented data, which is a key insight for making data augmentation truly effective in SEI.
Key Results
The TWC-CNN method demonstrated superior performance compared to six state-of-the-art SEI methods across various sample sizes on both ADS-B and WiFi datasets.
- ADS-B Dataset: With only 5% of training samples, TWC-CNN achieved an accuracy of 84.10%.
- WiFi Dataset: With only 5% of training samples, TWC-CNN achieved an accuracy of 96.99%.
- At 1% of the training samples, TWC-CNN achieved 42.90% on ADS-B and 52.00% on WiFi, significantly outperforming other methods which were below 35% and 20% respectively.
- T-SNE visualizations confirmed that TWC-CNN effectively separates features of different emitters and clusters features of the same emitter, showing better discriminative power than comparative methods.
Limitations and Future Work
- Computational Cost: The paper notes that the TWC-CNN method's training time is significantly increased due to the wavelet transform calculation and the use of two separate networks.
- Sample Imbalance: The authors acknowledge that even with sufficient samples, existing label distribution imbalances in datasets can bias models towards prevalent categories. They identify addressing sample imbalance as a focus for future work.
- CWT Scale Optimization: While the paper explores different CWT scales, it suggests that an increase in scale doesn't always improve performance and can even decrease it due to increased input data size. Further optimization of CWT parameters could be beneficial.
Summary updated