*********************************
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
*********************************
Title: An Ensemble Learning Approach to Differential Privacy for Deep Neural Network based Speech Processing
Committee:
Dr. Lee, Advisor
Dr. Heck, Chair
Dr. Moore
Dr. Chen
Abstract: The objective of the proposed research is to develop a new machine learning framework for differential privacy preserving speech processing. Speech signals contain a rich set of information that encompasses the characteristics of speakers and speaking environments. Differential privacy (DP) has recently emerged as a technique to provide identity protection with a mathematical definition for measuring privacy losses under query-based attacks under new data regulations (e.g., GDPR). However, ensuring DP in DNN models often leads to a severe performance loss due to the data perturbation process for achieving a strong privacy budget. This proposal aims to establish a perturbation-based ensemble learning scheme for training with DNNs to maintain the high performance of speech models while satisfying the new DP requirements. Our preliminary work used Laplace noise to guarantee epsilon privacy budget under a bounded max-divergence measurement with ensemble learning for speech recognition tasks. Furthermore, our proposed research will focus on theoretical studies of margin estimation to explore the tradeoff between ensemble model performances and DP. New privacy-preserving architectures with decentralized and ensemble training will also be investigated for the two above-mentioned speech applications.
