Abstract:Due to the diversity of RF signals, the complexity of the electromagnetic environment, and the difficulty of feature extraction, the robustness and applicability of the existing artificial features-based RF-specific emitter identification methods cannot meet the application requirements. Although the data-driven deep learning methods can provide a more flexible mode of specific emitter identification, they are less interpretable and lack a general test mode to evaluate their advantages and disadvantages. An evaluation method is explored for the deep learning model on the target individual dataset of the Electromagnetic Big Data Super Contest, and a general testing system architecture is proposed for the specific emitter identification model based on deep neural networks. The framework constructs the simulation test samples through signal feature masking, Generative Adversarial Network (GAN), deception signal collection, channel simulation and other methods, and imports the test samples and original data into the deep model to compare the recognition results. The test results are employed to judge the location of the signal key features extracted by the deep model, to analyze the robustness of the model, and to reveal the impact of the channel environment on the recognition performance, thus the performance of the deep learning model can be interpretable.