부산대학교 도서관

Applying deep learning (DL) to modulation recognition can significantly improve the efficiency of communication in cognitive radio (CR) systems, but it may be attacked by adversarial examples. The black-box attack has vital practical significance because it does not need to master the parameters and architecture of the target model. The ensemble attack is an essential black-box attack method, which attacks the model by improving the transferability of adversarial examples. However, the existing ensemble attacks only simply adopt the average method when fusing the outputs of different networks, without fully considering the characteristics of the ensemble model, resulting in poor transferability. This paper proposes an attention-based ensemble attack method, which uses the prediction performance of different networks to assign attention factors to express the influence of these networks, so that the example can pass through the decision boundaries of all networks within a limited number of iterations. Simulation results show that the proposed method can improve the transferability of adversarial examples and effectively attack the black-box modulation recognition model.