An Investigation into Deep Learning Based Network Intrusion Detection System for IoT Systems

Network Intrusion Detection System (NIDS) is a crucial part of the security implementation for service providers, organizations, and entities to protect connected devices on their networks. NIDS proactively detect and identify various threats and intrusions thus protecting sensitive data transmitted over the network. With the wide-scale adaptation and exponential growth of IoT devices connected to the network, having capable and highly efficient IDS becomes more necessary than ever. The NIDS rely on historical data to analyse the network information and to detect threats and intrusions. This paper conducts a comparative study of the use of modern deep learning models such as Fully Convolutional Network (FCN) and Autoencoder combined with Fully Connected Network (Autoencoder-FCN) to distinguish normal network data from attack data. This study uses CICIDS2017 dataset (over 2.8M network data records) representing real-world data and contains both normal data and attack data corresponding to the most up to date common attacks seen in modern network environments. The performance of both FCN and Autoencoder-FCN was observed to be highly accurate with the accuracy parameter being above 97% and low error rates. FCN model performs slightly better than the Autoencoder-FCN model. However, FCN model exhibited lower training time in a local deployment environment compared to the Autoencoder-FCN model.