Dissertation Talk: Taming Evasions in Machine Learning Based Detection Pipelines

Presentation: Departmental: CS | May 2 | 3-4 p.m. | 606 Soda Hall Speaker/Performer: Alex Kantchelian, UC Berkeley Sponsor: Electrical Engineering and Computer Sciences (EECS) Present day computing systems produce data at phenomenal rates, both in the form of user generated content or general cyber and physical measurements. Oftentimes, this data is used for driving security critical decisions at a quick pace. Due to the limited availability of human domain experts, Machine Learning (ML) techniques are the predominant enabler of those security decision pipelines. Such pipelines are, for example, responsible for detecting undesirable content in online social networks, assigning degrees of suspiciousness to large sets of unknown quality executables or recognizing critical environmental conditions in autonomous driving systems.In today’s talk, I will focus on a major conceptual flaw those pipelines share: their…


Link to Full Article: Dissertation Talk: Taming Evasions in Machine Learning Based Detection Pipelines