Distributed Machine Learning for IoT-based Networks

Abstract

Distributed Machine Learning has become common with the abundance of IoT devices, such as smart watches, smart thermostats, and many other sensors. This research aims at tackling the gaps in the research relating to distributed IoT-based networks where there exists a variety of experiences observed by the nodes. It also focuses on data privacy by constraining the types of information communicated between nodes. In this research, we devise novel collaborative learning methodologies for a connected network of edge nodes. In the approaches, class distribution information is shared to facilitate with the collaborative learning process. The first proposed methodology utilizes a diversity measure to compute the node weights. Whereas the second proposed methodology utilizes a similarity graph to compute the weights. The methodologies work by incentivizing the aggregated model to put an emphasis on nodes that have observed classes that are less common. The proposed methodologies are first developed using a synthetic dataset for fire sensitivity. Then, a case study using the PAMAP2 Physical Activity Monitoring dataset is implemented to validate the proposed approaches. A baseline method is identified, where node weights are computed using data form the performance of the models on the source nodes. To evaluate the performance of our proposed techniques, we used Leave-One-Subject-Out Cross Validation, comparing the accuracy, recall, and F1 score of all versions of our proposed approaches against the baseline model. The results show an improvement in all metrics for our proposed approaches compared to the baseline model. A significant result is the improvement in performance when class frequency information is shared between nodes, indicating the potential benefit of sharing higher resolution information in collaborative learning. This study serves as a foundation for future advancements in collaborative learning on edge networks, and encourages the development of scalable solutions that conform to data privacy constraints.

Date of Award	Aug 2023
Original language	American English
Supervisor	Deepak Puthal (Supervisor)