南京大学计算机科学与技术系

软件新技术与产业化协同创新中心

摘 要：

The past decade has seen the great potential of applying deep neural network (DNN) based software to safety-critical scenarios, such as autonomous driving. Similar to traditional software, DNNs could exhibit incorrect behaviours, caused by hidden defects, leading to severe accidents and losses. Hence, it is very important for quality and security assurance of deep learning systems, especially for those applied in safety- and mission-critical scenarios.

In this talk, I will introduce some of my latest research works in testing and security analysis on deep learning models. First, I will introduce DeepHunter (ISSTA'19), a coverage-guided fuzzing framework for testing feedforward neural networks. Second, I will introduce DeepStellar (FSE'19), a quantitive analysis for the stateful neural networks (e.g., RNN). At last, I will introduce DiffChaser (IJCAI'19), a differential testing framework for detecting disagreements which are caused by prediction on different models, frameworks or platforms.