A new universal camouflage attack algorithm for intelligent speech system

Abstract

The security problems in intelligent speech systems have been extensively investigated. In this paper, we propose a new more generic and more efficient attack method that can be used to promote the security of intelligent speech systems. By turning the attack target into a holistic intelligent speech system, we discovered security threats in speech resampling and risks in deep model normalization layers, and designed CA-ISS algorithms based on these two security risks. CA-ISS algorithm results in semantic content changes after speech resampling and launches an attack by creating cognitive differences between human and deep models. This paper also upgrades to the limitations of CA-ISS algorithm to attack the cloud platforms. Six intelligent speech systems are used to verify the effectiveness of CA-ISS in experiments. Experimental results demonstrate that the CA-ISS algorithm has sufficient generalisability, efficiency, and camouflage. Finally, the principle of CA-ISS algorithm is analyzed based on multiple visualization algorithms and the camouflage effect of the attack samples is evaluated.