University of North Dakota
Benchmarking Federated Learning Frameworks For Aviation Network Data
Abstract
dc:description.abstractAutomatic Dependent Surveillance-Broadcast (ADS-B) is an alternative technology adopted by the FAA instead of ground radar to enhance accurate navigation by relying on GPS satellites for precise aircraft position information. Factors such as jamming, multipath fading, and solar activities influence GPS data integrity issues, causing dropouts or missing data, thus affecting flight safety and navigational accuracy. To mitigate such potential GPS dropout-related incidents, there is a need for robust data-driven models. This thesis focuses on multiple studies: (1) investigate five distinct machine learning (ML) models to impute missing data on ADS-B/GPS information; (2) design a federated learning (FL) framework for aviation network data; and (3) conduct a benchmarking study to validate multiple quality attributes for the proposed aviation Fed-CPS framework. Preliminary results indicate (a) k-NN yields better accuracy over other ML models (Bayesian Ridge, Random Forest, AdaBoost, Extra Tree, and k-NN) even at the highest missing rate of 30%; (b) deployment of LSTM and k-Means in a federated setting indicate that LSTM results in both MAPE and computation run-time savings. Specifically, LSTM shows (i) performance-per-dollar of 1.5 times (client) and 0.5 times (server) than k-Means and (ii) energy-efficiency-per-watt of 1.5 times (client) and 0.5 times (server) than k-Means.
Degree
thesis:*- Name thesis:degree_name
- Master of Science (MS)
- Level thesis:degree_level
- Thesis
- Discipline thesis:degree_discipline
- Computer Science
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Subash Chandar, Barathwaja
- Contributors dc:contributor
-
- Prakash Ranganathan
Subjects
dc:subject × 1Identifiers
dc:identifier.*- Repository record dc:identifier
- https://commons.und.edu/theses/5705