Optimizing Distributed Machine Learning on User-Variant Edge Computing Systems

5G wireless networks and edge computing offer several advantages for mobile users, including reduced latencies and increased computing resources beyond local hardware constraints. Specifically, edge computing has a significant impact on distributed machine learning techniques such as federated learning and split learning, relying on distributed computational systems. These techniques allow model training on distributed data without sharing raw data and can benefit from reduced training and query times achieved through optimal allocation of training tasks. 

In this thesis, our objective is to optimize the training and testing of machine learning services in an edge computing system, while addressing challenges arising from the system’s heterogeneity. We must deal with the fact that each user has varying latency and resource requirements for tasks offloaded to the edge, and each edge server has limited and varying resource capacities. To tackle these issues, we first employ a combinatorial bandits framework to facilitate flexible resource identification and sharing. This approach is well-suited for reducing train time in a split learning scenario. Next, we formulate an integer linear program and its solution for the migration-based mobile edge computing problem. This solution allows deployed machine learning models associated with users to migrate, maintaining low latencies even when users are mobile. Our results indicate that migration-based policies significantly reduce the delay users experience when querying machine learning models in an edge computing environment. Finally, we investigate security threats due to variations in user intent, such as situations where certain users may attempt to sabotage the service for others by performing train time or test time attacks in a distributed machine learning setting.