Selective Audio Perturbations for Targeting Specific Phrases in Speech Recognition Systems

In this paper, we propose a novel approach for creating audio adversarial examples that specifically target speech recognition systems. The proposed method involves adding optimized noise to a particular region of an audio sample that corresponds to a specified word or phrase. By doing so, the generated adversarial example is designed to deceive the targeted model into interpreting it as a modified sentence, where the specified phrase has been altered. This method offers advantages compared to existing techniques, including reduced distortion since noise is only added to the targeted area, and the ability for the attacker to selectively modify or add specific words as desired. The experimental evaluation utilized the Mozilla Common Voice dataset. The results demonstrate that the adversarial examples generated using the proposed method, which only transform the specified word or phrase by adding noise to that specific region, can successfully fool the speech recognition system into misclassifying them as the intended target sentence.