A new universal camouflage attack algorithm for intelligent speech system

Dongzhu Rong, Qindong Sun, Yan Wang, Xiaoxiong Wang

Abstract: Highlights•We found security risks in speech resampling algorithms and deep model normalization layers of intelligent speech system.•A new camouflage attack algorithm CA-ISS for intelligent speech systems based on the two found security risks is proposed, and CA-ISS can attack the intelligent speech systems based on the cognitive differences between humans and deep models.•We improved CA-ISS algorithm for cloud platforms to make it more practical.•Experiments show CA-ISS algorithm has sufficient generalisability, it can effectively attack both intelligent speech recognition system and speaker recognition system.