Leveraging federated learning for DoS attack detection in IoT networks based on ensemble feature selection and deep learning models

Abstract

The Internet of Things (IoT) seamlessly integrates into daily life, enhancing decision-making and simplifying everyday tasks across various domains, including organizations, healthcare, the military, and industry. However, IoT systems face numerous security threats, making data protection against cyberattacks essential. While deploying an Intrusion Detection System (IDS) in a centralized framework can lead to data leakage, Federated Learning (FL) offers a privacy-preserving alternative by training models locally and transmitting only the updated model weights to a central server for aggregation. Detecting Denial-of-Service (DoS) attacks in IoT networks is critical for ensuring cybersecurity. This study compares the performance of centralized and federated learning (FL) approaches in detecting DoS attacks using four deep learning models: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Convolutional Neural Network (CNN). To enhance model efficiency, we apply filter-based feature selection techniques, including Variance Threshold, Mutual Information, Chi-square, ANOVA, and L1-based methods, and employ an ensemble feature selection approach by combining them through a union operation. Additionally, a wrapper-based Recursive Feature Elimination (RFE) method is used to refine feature selection by removing redundant and irrelevant features. Experiments were conducted using the IoT Intrusion Dataset (IoTID20), and model performance was evaluated based on accuracy, precision, recall, F1-score, and ROC-AUC metrics. In the centralized learning scenario, the highest accuracy was achieved with GRU using Mutual Information (MI) at 99.91 %, followed by RNN with MI at 99.90 %. In the FL scenario, the highest accuracy was achieved with CNN using the ANOVA method at 99.73 %, followed by GRU with Chi2 at 99.61 %. These findings underscore the significant impact of feature selection on learning performance and provide valuable insights into optimizing deep learning-based DoS detection in IoT networks.