Securing IoT Servers: Strategies for Employing Shallow and Deep Neural Networks

Abstract—This study investigates the potential employability of Shallow and Deep Feed Forward Neural Networks (FFNs) in detecting attacks on low-resourced IoT application servers. It employed a Shallow FFN model with a single hidden layer of 512 neurons and a Deep FFN model with 7 hidden layers, having between 256 to 4 neurons respectively. The study constructed four Shallow and Deep FFN models, utilizing two balanced UNSW-NB15 datasets containing 20 and 40 features. Experiments were conducted to detect network attacks on IoT networks. The results demonstrated that the Deep FFN model utilizing 40 features, despite slightly longer prediction times and higher resource usage, consistently outperformed other models, achieving an accuracy of 98.37%. Therefore, Deep FFN models prove suitable for protecting high-resourced IoT application servers. The Shallow model, achieving a faster detection time and moderate accuracy of 93%, is potentially employable in resource-constrained, low-latency IoT servers. This research enhances IoT security by employing Shallow and Deep FFN models based on different resource levels in IoT environments. Furthermore, it proposes integrating the Deep model into next-generation firewall systems to protect higher-value IoT servers. Future work involves exploring hybrid FFN architectures for protecting edge servers from network attacks.