Lightweight representation learning for network traffic towards malicious traffic detection in edge devices

With the rapid increase in the number of connected devices in the Internet of Things (IoT) environment, their exposure to threats has increased significantly. The attackers can launch sophisticated attacks on these networks more frequently due to the ease of availability of computing facilities. The devices in the IoT network have limited computational power, storage capacity, and hardware capability, making it challenging to secure them using traditional approaches. Over the years, many machine learning and deep learning-based approaches have been proposed to classify the traffic flowing through the edge devices, but the models have their limitations, such as slow detection of the attacks because of the limited computational power of these devices, thereby rendering parameter-heavy models infeasible to be run on such devices. To overcome this, we propose a structure learning algorithm to create a model whose structure learning is done using correlation analysis and PCA, then is optimized using parent divorcing and Sequential least squares programming, thereby creating a model that exhibits high performance despite being lean with respect to the number of parameters. The chosen features’ relevance for each attack is also validated via qualitative mapping and domain logic. The generated model, evaluated using UNSW-NB15 and TON-IoT datasets, outperformed several state-of-the-art models to classify malicious traffic, especially in terms of inference time and model size. Despite its resource efficiency, it shows comparable results in terms of accuracy, recall, precision, and F1 score with other baseline models.