SECURED FEDERATED LEARNING ENVIRONMENT DURING ADVERSARIAL ATTACKS

Abstract

Federated learning has emerged as a promising paradigm for collaborative machine learning across decentralized devices while ensuring data privacy. Federated Learning offers a novel and promising approach to machine learning, providing an alternative to traditional centralized methods, particularly beneficial in scenarios with challenges related to data confidentiality and quantity.Despite its advantages, the distributed nature of FL renders it susceptible to data poisoning attacks, where malicious clients can manipulate local data labels, leading to misclassifications by the global model. This thesis explores various advancements in federated learning aimed at overcoming key challenges and improving its practicality and effectiveness. While extensive research has been conducted on image datasets in federated learning, the performance and accuracy of FL on other types of data, such as CSV datasets, remain largely unexplored. To address this gap, this thesis focuses on using CSV datasets for loan default prediction to compare the performance of federated learning with centralized machine learning methods. Additionally, the impact of adversarial attacks, specifically label flipping attacks, on the performance of FL models will be investigated. To enhance the security and robustness of federated learning, this thesis proposes integrating homomorphic encryption techniques into the FL framework. By leveraging homomorphic encryption, the privacy of client data is preserved during the model training process, mitigating the risks associated with direct data sharing.The thesis will explore the performance of FL models on CSV datasets, comparing their accuracy and robustness to centralized machine learning methods. The impact of label flipping attacks will be analyzed, and the proposed integration of homomorphic encryption will be evaluated for its effectiveness in preserving privacy and mitigating the effects of adversarial attacks.