Specific Emitter Identification via Active Learning

TL;DR

The paper tackles the challenge of limited labeled data in specific emitter identification (SEI) by introducing an active-learning–driven, three-stage semi-supervised framework. It employs phase-rotation data augmentation and a self-supervised contrastive pretraining with a dynamic dictionary, followed by supervised fine-tuning that jointly optimizes cross-entropy and contrastive losses. An active sampling module using BALD (uncertainty) and K-center greedy (representativeness) selects informative unlabeled samples to label, reducing labeling costs. Experiments on ADS-B and WiFi datasets show substantial performance gains over traditional supervised and semi-supervised methods under low-label regimes, with the representativeness strategy often providing robust coverage in complex distributions. This approach offers a practical pathway to efficient data utilization and robust SEI in real-world wireless environments.

Abstract

With the rapid growth of wireless communications, specific emitter identification (SEI) is significant for communication security. However, its model training relies heavily on the large-scale labeled data, which are costly and time-consuming to obtain. To address this challenge, we propose an SEI approach enhanced by active learning (AL), which follows a three-stage semi-supervised training scheme. In the first stage, self-supervised contrastive learning is employed with a dynamic dictionary update mechanism to extract robust representations from large amounts of the unlabeled data. In the second stage, supervised training on a small labeled dataset is performed, where the contrastive and cross-entropy losses are jointly optimized to improve the feature separability and strengthen the classification boundaries. In the third stage, an AL module selects the most valuable samples from the unlabeled data for annotation based on the uncertainty and representativeness criteria, further enhancing generalization under limited labeling budgets. Experimental results on the ADS-B and WiFi datasets demonstrate that the proposed SEI approach significantly outperforms the conventional supervised and semi-supervised methods under limited annotation conditions, achieving higher recognition accuracy with lower labeling cost.

Figures (6)

• Figure 1: The block diagram of the three-stage training scheme.
• Figure 2: Neural network architectures in the encoder, the projection head, the predictor, and the classifier.
• Figure 3: The recognition accuracy of the ADS-B dataset is evaluated under different numbers of AL rounds. The initial number of labeled samples is $128$, and each round of active learning added $K=128$ newly labeled samples. For comparison, we also simulate a conventional CNN-based jian2020deep SEI approach and a semi-supervised SEI method based on CLwu2023specific.
• Figure 4: The recognition accuracy of the WiFi dataset is evaluated under different numbers of AL rounds. The initial number of labeled samples is $64$, and each round of active learning added $K=64$ newly labeled samples. The conventional CNN-based jian2020deep SEI approach and the CL SEI method wu2023specific are also simulated for comparison.
• Figure 5: Comparison of different AL selection strategies on a two-class dataset. (a) The uncertainty-based selection prefers samples near the decision boundary (green circles), which helps refine the classifier’s boundary but may over-focus on local hard cases. (b) The representativeness-based selection chooses globally diverse and informative samples (yellow stars), ensuring better coverage of the overall data distribution. The figure highlights the necessity of choosing appropriate AL strategies depending on dataset characteristics.