Enhancing Security: Federated Learning against Man-In-The-Middle Threats with Gradient Boosting Machines and LSTM

Abstract

Federated Learning (FL) can be preferred more in the cyber security community because it can enable training without sharing the actual training data when combined with different learning algorithms. Though these types of learning models maintain the privacy of the data by reducing the data centralization, it is vulnerable to cyber security threats like Man-In-The-Middle (MITM) attacks. The present paper makes an attempt to analyse the performance by integrating Gradient Boosting Machines (GBM) and Long Short-Term Memory (LSTM) with Principal Component Analysis (PCA) in a federated learning framework. Both the models are trained on a converted dataset using PCA that reduces the dimensionality of the data while maintaining its important values. The results indicated that the both models enhanced the confidentiality and integrity of the data against MITM attacks while FL-GBM performed better than FL-LSTM model in terms of accuracy and robustness in maintaining performance consistency across rounds of evaluations. This research further extends the current understanding of secure federated learning architectures and proposes a viable model for practical deployments in sensitive applications.