Improving Data Injection Attack Detection in Robotics Systems: A Comparative Study of Machine Learning Models

Abstract In recent decades, significant technological advancements have improved the integration of robots into environments shared with humans. However, these systems rely on conventional computing platforms, making them vulnerable to cyber threats that affect such platforms. Additionally, they present unique security challenges that can jeopardize privacy and physical safety. In this context, a new generation of robotics software has emerged, with the Robot Operating System (ROS) being one of the most widely adopted by researchers and developers. Nevertheless, various studies have shown that ROS is prone to vulnerabilities that could compromise its security and reliability. The present work proposes an improved, lightweight model based on network traffic anomaly analysis, while also comparing its performance against three well-known machine learning models: Random Forest, Naive Bayes, and a Fully Connected Neural Network. The experimental results obtained in a simulated realistic environment demonstrate that the proposed SVM-based approach achieved an accuracy rate of approximately 92% in detecting such attacks. Furthermore, the SVM model was compared to other learning models, such as Neural Networks, which showed promising results with slightly higher precision and specificity than the SVM.