A Deep Learning-Based Intrusion Detection System for Safeguarding ROS 2-Powered Unmanned Ground Vehicles Against DoS Attacks

Recent advances in robotics have increased the reliance on networked middleware, such as Robot Operating System 2 (ROS 2), for orchestrating Unmanned Ground Vehicles (UGVs) in dynamic and safety-critical settings. However, this distributed architecture exposes platforms like the Turtlebot4 to Denial-of-Service (DoS) vulnerabilities, which can incapacitate critical sensor and control messages. This paper presents a deep learning-based framework for near-real-time detection of DoS attacks in a ROS 2-enabled UGV. A novel adversarial script, generated using a Large Language Model (LLM), systematically floods targeted ROS 2 topics to simulate genuine network-based intrusions. The resulting dataset, comprising both benign and adversarial conditions, undergoes rigorous cleaning, feature engineering, and one-second binning to capture key metrics such as burstiness and topic entropy. Four neural architectures are evaluated: a Multilayer Perceptron (MLP), a Long Short-Term Memory (LSTM), a 1D Convolutional Neural Network (CNN), and an unsupervised Variational Autoencoder (VAE). Experimental findings show that the MLP, leveraging the engineered features, achieves 97% accuracy and an AUC of 0.998 in classifying DoS traffic, outperforming sequence-based models and the VAE. Integrated Gradients analysis further reveals that metrics reflecting abrupt message spikes and imbalanced topic usage are highly discriminative of malicious behavior. These results underscore the impact of domain-specific feature design in safeguarding ROS 2-based robotic platforms against emerging cyber threats.