Secure Distributed On-Device Learning Networks With Byzantine Adversaries

Due to the huge data volume in the big data era, the dimension of feature parameters and the size of datasets continue to increase. This trend induces a paradigm shift for the learning networks from the centralized in-cloud learning ones to the distributed on-device learning ones. Benefit from the parallel computation and local training, the distributed on-device learning networks have lower communication cost than the in-cloud learning networks. Moreover, the distributed on-device learning networks also have several good characteristics, such as scalability and privacy preserving. However, the distributed on-device learning networks are vulnerable to the malicious functions across the networks. The worse-case malicious functions are the Byzantine adversaries, where the malicious terminals compromise the learned model based on the full knowledge of the networks. Hence, the design of secure learning algorithms in the distributed on-device learning networks becomes an emerging topic. In the proposed research, we work on the development of a learning algorithm, which can secure the learning process and accelerate the convergence rate of the current learning algorithms.

Faculty Supervisor:

Victor C.M. Leung;Julian Cheng

Student:

Partner:

University of Minnesota (Twin Cities)

Discipline:

Engineering

Sector:

Information and Communications Technology; Technology

University:

The University of British Columbia

Program: