Abstract: Over the past decade, the application of learning-based software in domains like face recognition, autonomous driving, and content generation has showcased remarkable potential. As software evolves into intelligent systems, ensuring their quality and security becomes crucial, particularly in safety- and security-critical scenarios. However, analyzing and evaluating the quality of intelligent software presents significant challenges due to its black-box nature. In this talk, we address these challenges by introducing data-centric and model-centric analysis techniques for assessing the quality of deep learning-based software. We propose a label-free accuracy estimation method that enables evaluating the model performance on unlabelled data. Additionally, we develop an abstract model for deep neural networks, facilitating comprehensive analysis, testing, fault localization, and automated repair to enhance the software's quality and security.

Bio: Dr. Xiaofei Xie is an assistant professor at Singapore Management University. He obtained his Ph.D from Tianjin University and won the CCF Outstanding Doctoral Dissertation Award (2019) in China.  His research mainly focuses on the quality assurance of both traditional software and AI-enabled software. He has published some top-tier conference/journal papers in the areas of software engineering, security and AI,  such as ICSE, ESEC/FSE, ISSTA, ASE, TSE, TOSEM, ICLR, NeurIPS, ICML, TPAMI, Usenix Security and CCS.  In particular, he has received three ACM SIGSOFT Distinguished Paper Awards (FSE’16, ASE’19 and ISSTA’22).