Intrusion Detection System for Robot Operating System using Hybrid Deep learning

Robot operating system (RobotOS) has become an important standard middleware in industry and non-industrial sectors in modern world. Attacks to the RobotOS and its networks are increasing rapidly in recent days. These attacks lead to high risks and significant threats to the robot and its environment. This is because these robot systems are not designed by considering cyber security. Security issues of robot system receive increased attention by malicious attackers. This work presents robot network intrusion detection system (RobotNIDS) by leveraging the convolutional model and gated recurrent models in deep learning. The features of the specified deep learning models are flattened and combined. A low dimensional feature representation of the combined features is constructed by applying non-linear transformation with principal component analysis. Subsequently, two-phase classification model is applied on the low-dimensional features of network connection records to analyze the robot network activity in order to detect and classify attacks. The phase 1 employs KNN and random forest (RFort) classifiers and classification of predictions made in phase 1 is done using logistic regression (LRegn) in phase 2. Parameters of classification models are properly adjusted by considering the accuracy in hyperparameter experimental settings. Experiments on ROSIDS23 dataset shows that the proposed Robot-NIDS performs better than the existing intrusion detection and classification methods with an accuracy of $98.2 \%$ in detecting attacks and $98 \%$ accuracy in classification of attacks. The findings reported in this work demonstrates its ability for monitoring and analyzing robot network connection records to accurately detect the abnormal activity and report it to network administrator.