Abstract
Specific emitter identification (SEI) is a highly potential technology for physical layer authentication that is one of the most critical supplement for the upper-layer authentication. SEI is based on radio frequency (RF) features from circuit difference, rather than cryptography. These features are inherent characteristic of hardware circuits, which difficult to counterfeit. Recently, various deep learning (DL)-based conventional SEI methods have been proposed, and achieved advanced performances. However, these methods are proposed for close-set scenarios with massive RF signal samples for training, and they generally have poor performance under the condition of limited training samples. Thus, we focus on few-shot SEI (FS-SEI) for aircraft identification via automatic dependent surveillance-broadcast (ADS-B) signals, and a novel FS-SEI method is proposed, based on deep metric ensemble learning (DMEL). Specifically, the proposed method consists of feature embedding and classification. The former is based on metric learning with complex-valued convolutional neural network (CVCNN) for extracting discriminative features with compact intra-category distance and separable inter-category distance, while the latter is realized by an ensemble classifier. Simulation results show that if the number of samples per category is more than 5, the average accuracy of our proposed method is higher than 98%. Moreover, feature visualization demonstrates the advantages of our proposed method in both discriminability and generalization.
1. Introduction: The Challenge of Data Hunger in SEI
Specific Emitter Identification (SEI) serves as a "physical layer" security technique that identifies a specific piece of hardware based on unique, unintentional distortions in its radio frequency (RF) output—essentially a hardware fingerprint. Unlike cryptographic keys, these fingerprints are nearly impossible to counterfeit because they stem from microscopic manufacturing tolerances in analog circuits.
However, traditional deep learning models for SEI face two major hurdles: Data Hunger and Close-set Limitations. Most existing methods require thousands of samples per device to learn effectively and struggle to identify "new" devices that weren't part of the original training set. In real-world scenarios, such as monitoring aircraft via Automatic Dependent Surveillance-Broadcast (ADS-B) signals, we often encounter new aircraft for which we only have a few signal bursts. This paper addresses the Few-Shot challenge: how to build a system that generalizes to unseen emitters using extremely limited data.
2. Proposed Methodology: Deep Metric Ensemble Learning (DMEL)
The researchers propose a Deep Metric Ensemble Learning (DMEL) framework, which splits the task into two distinct phases:
2.1 Offline Feature Embedding
The system utilizes a Complex-Valued Convolutional Neural Network (CVCNN). Since RF signals are naturally represented as complex IQ (In-phase and Quadrature) data, CVCNNs are mathematically better suited to capture phase and amplitude relationships than standard real-valued networks. To ensure the features are discriminative (tight clusters for the same device, wide gaps between different devices), the authors employ a Hybrid Metric Loss combining Softmax Loss, Triplet Loss, and Center Loss.
2.2 Online Ensemble Classification
Once the CVCNN is trained to extract high-quality features, a simple, lightweight classifier (like Logistic Regression) is used for the actual few-shot task. To boost stability and accuracy, an Ensemble Learning approach is used, averaging the probability outputs of multiple independent base models to reduce the impact of noisy or low-quality signal samples.
3. Experimental Setup and Datasets
The study utilizes a dataset of ADS-B signals collected from real-world aircraft, divided into two parts:
- Auxiliary Dataset ($D_{au}$): Contains signals from 90 aircraft with 200–500 samples each, used to pre-train the CVCNN.
- Few-Shot Dataset ($D_{fs}$): Contains signals from "new" aircraft not seen during pre-training. The models were tested on 10, 20, and 30 "ways" (number of aircraft to distinguish) and 1 to 20 "shots" (samples available per aircraft).
4. Key Results and Insights
- High Accuracy: In a "10-way" scenario, the model achieved 98.2% accuracy with just 5 samples (5-shot) and nearly 99% with 20 samples.
- Separability vs. Discriminability: Standard deep learning focuses on separability. This paper demonstrates that for few-shot learning, features must be discriminative—so distinct and compact that even a simple linear classifier can distinguish new classes.
- Stability via Ensembles: Few-shot learning is sensitive to "bad" samples (e.g., signals captured during a momentary fade). The ensemble approach effectively "votes out" errors caused by these outliers, providing more robust performance.
5. Conclusion and Future Work
While the proposed DMEL framework achieves high average accuracy, performance can still fluctuate depending on the specific few-shot samples chosen (ranging from 69% to 98% in 1-shot scenarios). Future research will focus on integrating Attention Mechanisms to help the network focus on the most stable parts of the RF signal, potentially mitigating the impact of noise and further stabilizing performance in extreme low-data scenarios.
Summary updated