Federated Learning-Based Network Intrusion Detection Using Deep Auto Encoder

Abstract

The progress of information technology is directly related to different elements of daily life, providing people with services that make living more comfortable. As the network infrastructure grows to support new services, it necessarily produces various weak points that cybercriminals might exploit. Building a reliable network intrusion detection system with deep learning methodologies demands a substantial dataset. In the past, the conventional approach involved collecting data for centralized learning, which was then transmitted to a central server for model training. Nevertheless, this approach raises valid concerns regarding the potential exposure of personal information contained within raw data, which could lead to legal repercussions for data providers. Consequently, this study presents a Modified FedAvg algorithm to enhance the existing FedAvg algorithm for network intrusion detection models. This modified approach maximizes the potential of federated learning, enhancing data privacy and minimizing the transmission of model weights. Employing a Deep Auto Encoder within a federated setting, we evaluated our methodology using the CICIDDoS 2019 dataset, attaining a notable accuracy rate of 98.82% in effectively identifying intrusions.