Adversarial Attack Based-on Transferability for Specific Emitter Identification

Deep learning model for specific emitters identification(SEI) is vulnerable to adversarial samples generated by the application of subtle perturbations. However, it is observed that gradient-based adversarial samples generated by white box are poorly effective when they transfer to attack black box. This situation creates the misleading impression that the SEI model possesses excellent robustness. In this study, we propose Virtual Model Gradient Attack (VGA) to improve the attack success rates in black box and enhance the transferability. Firstly, virtual model endows the white box with multiple decision boundaries, aiming to extract multiple gradient information from a single model. Then, VGA achieves fast convergence of the gradients and avoids the occurrence of local optima during the convergence process through the implementation of a gradient update policy. Finally, the simulation results demonstrate that VGA outperforms the existing transferability-based attack algorithms, and the adversarial samples generated by VGA has excellent concealment.