Federated Learning in Email Server Authentication for Spam Mail Detection Using LSTM Algorithm

To propose a privacy-preserving spam detection system based on federated learning with LSTM that facilitates decentralized training while improving model performance collectively. Materials and methods: The system employs publicly available spam email datasets like Enron and Spam Assassin, a combination of spam and non-spam emails. Conventional spam detection systems are based on centralized machine learning models. The system employs federated learning with LSTM, locally training on the user’s device without data sharing. Federated Learning (FL) performs better than Centralized Learning (CL) in accuracy ($94.5 \%$ vs. $91.0 \%$), precision, recall, and F1-score, and detection performance. Statistical analysis and T-test results confirm FL’s significant improvements, including quicker detection time (0.759s vs. 1.210s) and better precision, recall, and F1-score, and making it a better and efficient spam detection tool with significance of 0.05. Spam detection from email systems based on federated learning and LSTM is secure and fast.