Boosting Al Speech Adversarial Example Transferability Via Time-Frequency Smoothing

Deep neural networks exhibit susceptibility to adversarial perturbations, wherein minimal modifications can drastically impact their predictions. Although the field of image-based adversarial attacks has been extensively explored, attacks in the speech domain are more complex due to the dynamic nature and rich frequency information of speech signals. This paper proposes a novel adversarial example generation method based on timefrequency smoothing to improve transferability against speech anti-spoofing models. By applying time-frequency smoothing during optimization, our approach enhances adversarial perturbations and prevents them from being trapped in local optima. To further enhance performance, we introduce an ensemble strategy that combines two distinct models to improve the success rate of black-box attacks. Extensive experiments demonstrate that our time-frequency smoothing method achieves superior results in both white-box and black-box settings, with performance improvements ranging from 1.81 % to 8.29 % compared to state-of-the-art (SOTA) methods. Additionally, the two-model ensemble attack enhances performance further, achieving improvements ranging from 0.17 % to 7.38 % over SOTA methods.