A Decomposed Deep Training Solution for Fog Computing Platforms

Abstract

Legacy machine learning solutions collect user data from data sources and place computation tasks in the Cloud. Such solutions eat communication capacity and compromise privacy with possible sensitive user data leakage. These concerns are resolved by Fog computing that integrates computation and communication in Fog nodes at the edge of the network enabling and pushing intelligence closer to the machines and devices. However, pushing computational tasks to the edge of the network requires high-end Fog nodes with powerful computation resources. This paper proposes a method whose computation tasks are decomposed and distributed among all the available resources. The more resource-demanding computation is placed in the Cloud, and the remainder is mapped to the Fog nodes using migration mechanisms in Fog computing platforms. Our presented method makes use of all available resources in a Fog computing platform while protecting user privacy. Furthermore, the proposed method optimizes the network traffic such that the high-critical applications running on the Fog nodes are not negatively impacted. We have implemented the (deep) neural networks - using our proposed method and evaluated the method on MNIST and CIFAR100 as the data source for the test cases. The results show advantages of our proposed method comparing to other methods, i.e., Cloud computing and Federated Learning, with better data protection and resource utilization.