Blind adversarial attack based on time-frequency models for speech recognition systems

As intelligent devices become increasingly prevalent in our daily life, the requirement of privacy has been significantly increased. To address the issue of privacy protection, the topic of adversarial attack appeared in recent year. Initially, adversarial attack was predominantly applied to image recognition. However, due to the unique characteristics of audio data, the attacks suitable for images, e.g., additive perturbations, may not be applicable in audio cases. The goal of this study is to perform adversarial attack on speech signals such that they cannot be recognized by automatic speech recognition (ASR) systems but still be identified by humans. We introduce several distinct methods for noise addition and precision-reducing to generate adversarial examples for ASR systems. The proposed approach leverages audio features extracted through filtering and time-frequency transformations. The adversarial samples generated using the proposed methods not only retain their intelligibility for human listeners but also achieve a 100% success rate in blind attacks against ASR systems with unknown architectures and parameters.