Adversarial Attack Generation Based on Meta Learning in Specific Emitter Identification

Abstract

Specific emitter identification (SEI) based on deep learning is a highly potential physical layer security authentication technology. However, deep neural networks (DNNs) are vulnerable to adversarial examples. In this letter, we propose an adversarial attack method based on meta-learning strategies (Meta-SA) against SEI. Meta-SA according meta-learning idea to explore the dynamic decision-making of attack perturbation parameters. It can quickly identify the characteristics of the SEI model and evolve defense mechanisms, thus the optimal attack parameters are adjusted to generate new effective attacks. We define the meta model to adjust the attack parameters to maximize the effect of the attack, enhance the attack covertness and improve the overall success rate of the attack. In order to verify the feasibility of Meta-SA, experiments are based on the actual collected signal ADS-B dataset, the results show that Meta-SA has good performance and the high recognition model ResNet accuracy is minimized to 9.51%.