Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/FA6C4431-605B-4AD9-96CE-A8974BF1B006 | ||
| Year | 2024 | ||
| Type of Item | Diploma Work | ||
| License |
|
||
| Bibliographic Citation | Panagiotis Sklavos, "Federated learning at flower and NS3 integrating the geometric approach for efficient synchronization", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2024 https://doi.org/10.26233/heallink.tuc.100437 | ||
| Appears in Collections |
The rapid advancement of edge computing and the Internet of Things (IoT) has led to an exponential increase in data generation, underscoring the need for privacy-preserving and efficient decentralized machine learning methods. This thesis addresses these needs by implementing Federated Learning (FL) under realistic network conditions, integrating the Flower Framework with the NS3 network simulator, and employing the Geometric Method (GM) for enhanced synchronization and performance. Our approach involves several key steps. Initially, a Federated Learning orchestrator is developed using the Flower framework to establish a distributed network with a central server and multiple clients connected in a star topology. To optimize model updates andminimize communication costs, a synchronization method based on geometric monitoring, known as Functional Dynamic Averaging (FDA), is implemented. Additionally, the NS3 network simulator is used to emulate realistic network conditions, and a socket-based Inter-Process Communication (IPC) protocol is employed to ensure seamless interaction between the federated learning framework and the network simulator. Our integrated FL framework demonstrates robustness and effective synchronization across various simulated network conditions. The Geometric Monitoring approach of FDA efficiently balances the computation-to-communication ratio while maintaining high accuracy levels. Thorough testing across diverse datasets, artificial neural networks (ANNs),networking conditions, and data distributions (IID and non-IID) reveals significant improvements in communication efficiency and model accuracy compared to a baseline distributed algorithm. In conclusion, this research presents a novel and effective Federated Learning framework that bridges existing infrastructure gaps, ensuring robust performance and efficient synchronization in real-world network environments.
Copyright © DIAS 2013
