A Robust Defence Strategy for Secure Federated Learning Using LOMAR-Based Validation

Abstract

This project presents a secure and efficient federated learning framework designed to improve model accuracy while defending against malicious updates. The system enables a centralized server and multiple clients to collaboratively train machine learning models without sharing raw data. A graphical interface is used to upload and preprocess the MNIST dataset, train models, and communicate with the server. The proposed approach incorporates LOMAR-based techniques to evaluate model integrity and detect poisoned updates. Experimental results show that genuine models achieve high accuracy, while malicious models are identified and rejected by the server. Additionally, model compression techniques are applied to reduce communication overhead, improving efficiency in distributed environments. Comparative graphs demonstrate that the proposed method outperforms traditional approaches in terms of accuracy and model size. Overall, the system ensures secure, scalable, and reliable federated learning suitable for real-world applications.