Enhancing DDoS Detection in SDN-Based Autonomous Vehicles using Transformer-based Feature Transformation

Abstract

Autonomous vehicles (AVs) require fast, reliable, and secure communication to function safely in dynamic environments. To meet these demands, Software-Defined Networking (SDN) is increasingly used in AV systems due to its centralized control and ability to manage network traffic efficiently. However, this centralization also makes SDN vulnerable to Distributed Denial-of-Service (DDoS) attacks, which can disrupt critical vehicle operations. While earlier studies have used Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) for DDoS detection, this work introduces a Transformer-based feature selection method that leverages self-attention to focus on the most relevant network traffic features. This improves both detection accuracy and response time. To further enhance performance, we apply Bayesian Optimization with HyperNet (BOHN) for tuning model parameters. For interpretability, SHapley Additive exPlanations (SHAP) is used to highlight the impact of each feature on the model's predictions. The model is trained on the InSDN dataset, which simulates real-world SDN attack scenarios, and tested on the UNSW-NB15 dataset, a comprehensive benchmark containing a mix of normal and malicious traffic. Experimental results show significant improvements in accuracy, false positive reduction, and adaptability to new types of threats, demonstrating the value of self-attention-based approaches for securing SDN-enabled AV networks.