Mahembe, Bright Kudzaishe2025-10-292025-10-292025https://hdl.handle.net/10566/21327Enhancing federated learning performance in realistic network conditions using a customised udp protocol and greedy parameter aggregation Federated Learning (FL) is an emerging paradigm enabling decentralized Machine Learning (ML) model training and updates while prioritizing data privacy. Extensive research in this field has led to the development and enhancement of multiple solutions, including as TensorFlow and PyTorch, to facilitate decentralized ML simulations. However, existing frameworks often have limited capabilities for customizing network configurations and transport protocols to create realistic network environments. Transport layer protocols, situated at Layer 4 of the Open Systems Interconnection (OSI) model, facilitate end-to-end communication between multiple hosts, ensuring data is reliably transmitted across the network. This study leverages the NS-3 network simulator alongside Ten- Sor Flow's deep learning framework to create a realistic network environment tailored for Federated Learning applications. To address the specific efficiency and reliability requirements of these applications, a modified User Datagram Protocol (UDP) was developed. A detailed implementation of the proposed NS-3-based Federated Learning simulator is provided, along with an in-depth explanation of the modified UDP protocol. The simulator was employed to validate the simulation by comparing the performance of standard and modified UDP protocols using the CIFAR-10 (Canadian Institute for Advanced Research) and MNIST (Modified National Institute of Standards and Technology) datasets. Results indicate that the modified UDP model demonstrated robust performance, achieving an accuracy of 78% under poor network conditions, representing only a 2% decline from the 80% accuracy attained in ideal conditions. This performance is primarily attributed to its effective packet retrieval mechanism, whereas the standard UDP protocol model suffered a significant performance drop, achieving only 10% accuracy under poor network conditions, corresponding to a 69% decline from its performance in ideal conditions.enFederated LearningMachine LearningArtificial IntelligenceCloud ComputingEmerging ParadigmThesis