Acceleration and Privacy Protection for Distributed Computing

Abstract

In recent years, the rapidly increasing data base scales and complexities arising from various engineering, social and biomedical applications strongly motivate the study of distributed computation. The central idea is that a group of decentralized agents can break down and solve complex problems with suitable interconnection and cooperation, in which one gains scalability and resilience compared to traditional centralized computations. New challenges arrive accordingly in terms of the development and analysis of distributed algorithms, the trade-offs between communication complexity and computation efficiency, and the awareness of security and privacy in a distributed scheme. The seminar presents our contributions to this field in an effort to systematically tackle those challenges. 1. Novel distributed algorithms in light of the seminal Arrow-Hurwicz-Uzawa flow and stochastic gradient optimization are developed, for solving least-squares problems over networks, which is arguably one of the most important computation tasks. 2. A simple yet powerful approach for the acceleration of the conventional gossip protocols, which have become the canonical solutions for distributed information dissemination, is proposed by exploring local clique structures. 3. The systematic privacy leakage risks in the existing network linear equation solvers are uncovered rigorously. Universal privacy-preserving algorithms, with mathematically proven privacy protection guarantee, are also designed, which can be used as a data-encryption subroutine in a variety of distributed algorithms. These results add to the fundamental understandings of distributed computing architecture in emerging applications such as smart grid, intelligent transportation, and social computing, etc.