SkillExplorer: Understanding the Behavior of Skills in Large Scale
Zhixiu Guo, Zijin Lin, Pan Li, and Kai Chen, SKLOIS, Institute of Information Engineering, Chinese Academy of Sciences, China; School of Cyber Security, University of Chinese Academy of Sciences, China
Smart speakers have been popularly used around the world recently, mainly due to the convenience brought from the virtual personal assistant (VPA) which offers interactive actions through the convenient voice commands from users. Besides the built-in capabilities, VPA services can be further extended by third-party developers through skills. Similar to smartphone applications on Android and iOS markets, skills are also available on markets (e.g., Amazon, Google), attracting users together with malicious developers. Recent researches discover that malicious developers are able to route users' requests to malicious skills without users' consent by creating skills with similar names of legitimate ones. However, to the best of our knowledge, there is no prior research that systematically explores the interaction behaviors of skills, mainly due to the challenges in handling skills' inputs/outputs which are in the form of natural languages. In this paper, we propose the first systematic study on behaviors of skills, which is achieved by a suite of new grammar-based techniques including utterance extraction, question understanding, and answer generation specifically designed for skills. We build an interactive system called SkillExplorer and analyze 28,904 skills from the Amazon market and 1,897 actions from the Google market. Among these skills, we find that 1,141 skills request users' private information without following developer specifications, which are actually demanded by markets. 68 skills continue to eavesdrop users' private conversations, even after users have sent the command to stop them.
View the full USENIX Security '20 program at https://www.usenix.org/conference/usenixsecurity20/technical-sessions