Defending against audio adversarial examples based on multiple-sub-detectors
Keiichi Tamura, Hajime Ito · International Journal of Computational Intelligence Studies · 2022
Audio adversarial examples are audio input that can deceive the speech-to-text (STT) transcription neural network technology. In this paper, we propose a new defence method based on the decision-making of multiple sub-detectors to ensure security against adversarial examples for STT transcription neural networks. This method utilises a detector for defending audio adversarial examples consisting of three different sub-detectors. Then the results of the three sub-detectors are used to make the final decision, which makes the model more robust and accurate than a single detection method. Experiments were conducted to evaluate the method using 2,000 voice sound data, including 1,000 normal voice sound data and 1,000 audio adversarial examples, created on Mozilla-implementation DeepSpeech. The results of experiments to detect audio adversarial examples confirmed that the proposed method offers better performance in protection than our previous method.