ABSTRACT: Machine learning and deep learning in particular have become common tools in systems and security research. Aside from inherent problems with interpretability, the impact of deep learning has been largely positive. In this talk, I discuss some of the less studied problems and challenges of deep learning. First, I will talk about the threats posed by recent advances in natural language processing and DNNs, in the context of automated systems for malicious crowdsourcing. Today’s “crowdturfing” attacks are limited by costs of hiring and managing workers. Our work identified automated attacks that leverage deep learning language models to automate the generation of fake online reviews for products and services. Using Yelp reviews as a platform, we show how a customized NLP model can produce reviews indistinguishable from user-written reviews by state-of-the-art software tools. We use a user study to show that they are also undetected by human readers, and also score high on “usefulness” metrics. We develop novel defenses that identify the lossy transformation introduced by the RNN training and generation cycle.  If time permits, I will also briefly summarize some of our upcoming work looking at the performance vs. complexity tradeoffs of black box ML classifiers. Expanded versions of these results will be presented later this year in presentations at ACM CCS and IMC, respectively. Finally, I’ll talk briefly about the recent growth in CS at University of Chicago, and the new focus on systems and data, and interdisciplinary projects.

 

BIOGRAPHY: Ben Y. Zhao ( 赵燕斌 ) is the Neubauer Professor of Computer Science at University of Chicago.  He completed his PhD from Berkeley (2004) and his BS from Yale (1997). He is an ACM distinguished scientist, and recipient of the NSF CAREER award, MIT Technology Review's TR-35 Award (Young Innovators Under 35), Computer World Magazine's Top 40 Tech Innovators award, Google Faculty award, and IEEE ITC Early Career Award. His work has been covered by media outlets such as New York Times, Boston Globe, Fortune, Forbes, Scientific American, LA Times, MIT Tech Review, and Slashdot. He has published more than 150 publications in areas of security and privacy, networked systems, wireless networks, data-mining and HCI (H-index 58). He recently served as TPC co-chair for the World Wide Web Conference (WWW 2016), and is an associate editor for ACM Transactions on the Web.