A Speech Adversarial Sample Detection Method Based on Manifold Learning

Deep learning-based models have achieved impressive results across various practical fields. However, these models are susceptible to attacks. Recent research has demonstrated that adversarial samples can significantly decrease the accuracy of deep learning models. This susceptibility poses considerable challenges for their use in security applications. Various methods have been developed to enhance model robustness by training with more effective and generalized adversarial examples. However, these approaches tend to compromise model accuracy. Currently proposed detection methods mainly focus on speech adversarial samples generated by specified white-box attack models. In this study, leveraging manifold learning technology, a method is proposed to detect whether a speech input is an adversarial sample before feeding it into the recognition model. The method is designed to detect speech adversarial samples generated by black-box attack models and achieves a detection success rate of 84.73%. It identifies the low-dimensional manifold of training samples and measures the distance of a sample under investigation to this manifold to determine its adversarial nature. This technique enables the preprocessing detection of adversarial audio samples before their introduction into the deep learning model, thereby preventing adversarial attacks without affecting model robustness.