ISCA Archive Interspeech 2025
ISCA Archive Interspeech 2025

CBA: Backdoor Attack on Deep Speech Classification via Audio Compression

Yuheng Huang, Ying Ren, Wenjie Zhang, Diqun Yan

As deep neural networks permeate various aspects of society, research has demonstrated their vulnerability to backdoor attacks, particularly with untrustworthy third-party platforms, enabling attackers to manipulate model outputs using customized triggers. However, most audio-based backdoor methods are either sample-agnostic or audible. In this paper, we propose Compression-based Backdoor Attack (CBA), a novel strategy designed specifically for audio compression scenarios, which also utilizes fusion as well as iterative optimization strategies to generate sample-specific triggers and ensure both the effectiveness and stealthiness of the triggers. Extensive experiments validate the attack efficiency and robustness.