Enabling Fast and Universal Audio Adversarial Attack Using Generative Model

• DOI: 10.1609/aaai.v35i16.17663
• Publication Date: 2022-09-08T20:14:09Z
• Authors: Yi Xie; Zhuohang Li; Cong Shi; Jian Liu; Yingying Chen; Bo Yuan
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Sound (cs.SD); Audio and Speech Processing (eess.AS); FOS: Electrical engineering, electronic engineering, information engineering; 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology; Computer Science - Sound; Electrical Engineering and Systems Science - Audio and Speech Processing; Machine Learning (cs.LG)

Recently, the vulnerability of deep neural network (DNN)-based audio systems to adversarial attacks has obtained increasing attention. However, existing allow adversary possess entire user's input as well granting sufficient time budget generate perturbations. These idealized assumptions, however, make mostly impossible be launched in a timely fashion practice (e.g., playing unnoticeable perturbations along with streaming input). To overcome these limitations, this paper we propose fast perturbation generator (FAPG), which uses generative model for single forward pass, thereby drastically improving generation speed. Built on top FAPG, further universal (UAPG), scheme craft that can imposed arbitrary benign cause misclassification. Extensive experiments DNN-based show our proposed FAPG achieve high success rate up 214X speedup over attack methods. Also UAPG generates much better performance than state-of-the-art solutions.