Add to Google Calendar
Date:  Fri, May 06, 2016
Time:  10:00 am - 11:00 am
Location:  Holmes Hall 389
Speaker:  Dr. Yao Zheng, Ph.D. candidate, Virginia Tech

Abstract

Deep learning with artificial neural networks has become the new favorite for Big Data Analytics. By learning multiple levels of representations from the input data, deep learning methods can achieve unprecedented accuracy in a variety of information processing tasks. However, the massive volume of data required to successfully train a deep neural network also raises privacy concerns. Specifically, centrally stored data collection is often the targets of various malicious behaviors and subjects to legal subpoenas and extra-judicial surveillance. The situation becomes direr when the data contains personally identifiable information such as face images, voices, and other biometrics. The privacy and confidentiality concerns may prevent data curators from sharing data and thus benefitting from large-scale deep learning. To apply deep learning methods to sensitive data, we designed a distributed training method that allows multiple participants to collectively learn a deep neural network, without violating the privacy of individual data records. By interleaving the Laplace mechanism into different training phases of deep learning, our method can ensure differential privacy of the participants’ datasets, while still exploring the data diversity to boost the neural network’s overall performance.

Biography

Yao Zheng is a Ph.D. candidate at Virginia Tech. He received his B.S. degree in Microelectronic from Fudan University in 2007, M.S. degree in Electrical Engineering from Worcester Polytechnic Institute in 2011. Between 2007 and 2009, he was a Researcher at Siemens AG, Beijing, China. Yao’s research area is information security and privacy, machine learning and cryptography.