Generating Adversarial Examples in Audio Classification with Generative Adversarial Network
Qiang Zhang, Jibin Yang, Xiongwei Zhang, Tieyong Cao · 2022 7th International Conference on Image, Vision and Computing (ICIVC) · 2022
To improve the performance of acoustic adversarial examples, this paper proposes an adversarial generation model based on Generative Adversarial Network (GAN) for audio classification. By introducing the classification model into GAN, this paper proposes a general GAN framework to execute adversarial attacks for audio classification. Then we propose a Short-time Synthesis GAN-based (SSGAN) attack method, which can reduce the complexity of audio adversarial example generation, and further improve the generality and performance of the GAN-based audio adversarial example generation. Experiments on audio classification datasets such as UrbanSound8k and ESC50 show that compared with existing audio adversarial example generation methods, the proposed method generates adversarial examples with lower perceptibility, and has a higher attack success rate and attack efficiency for typical audio classification models.