Software is powering every aspect of our society, but it remains plagued with errors and prone to critical failures and security breaches. Program analysis has been a predominant technique for building trustworthy software. However, traditional approaches rely on hand-curated rules tailored for specific analysis tasks and thus require significant manual effort to tune for different applications. While recent machine learning-based approaches have shown some early promise, they, too, tend to learn spurious features and overfit to specific tasks without understanding the underlying program semantics. 

 

In this talk, I will describe my research on building machine learning (ML) models toward learning program semantics so they can remain robust against transformations in program syntax and generalize to various program analysis tasks and security applications. The corresponding research tools, such as XDA, Trex, StateFormer, and NeuDep, have outperformed commercial tools and prior arts by up to 117x in speed and by 35% in precision and have helped identify security vulnerabilities in real-world firmware that run on billions of devices. To ensure the developed ML models are robust and generalizable, I will briefly describe my research on building testing and verification frameworks for checking the safety properties of deep learning systems. The corresponding research tools, such as DeepXplore, DeepTest, ReluVal, and Neurify, have been adopted and followed up by the industry (e.g., in TensorFuzz built by Google), been covered in media such as Scientific American, IEEE Spectrum, Newsweek, and TechRadar, and inspired over thousands of follow-up projects.