Electronic Control Unit (ECU) Identification for Controller Area Networks (CAN) using Machine Learning

Electronic control units (ECUs) play a critical role in passengers’ safety and maintaining optimal vehicular performance. They control several vehicular functionalities such as adjusting seat positions or driver/passenger windows and wiper and headlight control. Identifying these ECUs is a complex, tedious process and requires extensive reverse engineering of the control area network (CAN) data set, as a modern car may have anywhere from 40 to 150 ECUs. Thus, individual ECU identification can be made at scale by leveraging data-driven analysis and machine learning by exploiting their idle and driving state conditions patterns. In this paper, three machine algorithms (k-nearest neighbors, Gaussian Naive Bayes, and Decision Tree) were analyzed to classify five different ECU signatures (e.g., braking, steering, speed, tachometer, and lighting) from 3 vehicles make (Honda, Nissan, and Toyota). The kNN model showed accurate classification performance with average accuracy and F1-scores of 94.4% and 0.94, respectively, followed by the Decision Tree and the Gaussian Naive Bayes models. Results from the cross-validation check demonstrate that the Decision Tree and k-nearest neighbors models can generalize well based on three feature inputs (time, size, and data payload) from CAN data.