*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Tess Malone, Communications Officer
Summary Sentence: Transparent Computing Systems Enabled by Program Analysis
Full Summary: No summary paragraph submitted.
Shiqing MaTITLE: Transparent Computing Systems Enabled by Program Analysis
ABSTRACT:
Modern computing systems are complex and opaque, which is the root cause of many security and software engineering problems. In enterprise level system operations, this leads to inaccurate and hard-to-understand attack forensics results, and significant runtime and space overhead. In deep learning systems, such opaqueness prevents us from developing scientific ways to improve the trained models and combating adversarial sample attacks. Hence, there is a pressing need for improving the transparency of these systems to help us solve the corresponding security and software engineering problems.
In this talk, I will focus on my research efforts of developing novel program analysis techniques to improve the transparency of such systems and their applications in attack forensics and deep learning systems. For attack forensics, I will first describe an annotation-based execution partitioning technique MPI which helps accomplish accurate, semantics-rich and multi-perspective attack forensics. Then I will introduce my novel provenance tracking system design which leverages the accurate analysis results enabled by execution partitioning to achieve low runtime and space overhead by performing online system event redundancy detection and reduction. For deep learning systems, I will discuss how two widely used software engineering techniques, state differential analysis and input selection, are leveraged to analyze deep learning model internals for addressing underfitting and overfitting problems; and how deep learning model invariants that are analogous to program invariants can be used to defend against adversarial sample attacks Finally, I will briefly present my ongoing and future work on intelligent system (i.e., systems that combine traditional computing components and artificial intelligent components) security and productivity.
BIO:
Shiqing Ma is a Ph.D. candidate in the department of computer science at Purdue University, co-advised by Professors Xiangyu Zhang and Dongyan Xu. His research interests lie in solving security and software engineering problems via program analysis techniques with a focus on improving the transparency of modern computing systems. He is the recipient of two Distinguished Paper Awards at ISOC NDSS 2016 and USENIX Security 2017.
