Semi-supervised intrusion detection for resource-constrained vehicular networks

To address issues in traditional in-vehicle network intrusion detection, such as need for large amount of labeled data and difficulty to detect unknown attacks, a lightweight intrusion detection method based on semi-supervised learning and dense pseudo-labeling is proposed． Three processes are involved in this new method． Curriculum-based labeling algorithm is used to assign pseudo-labels to unlabeled data． Original labels and LAN identifier fields of data with pseudo-labels are binary encoded and converted into images, a dense convolutional network is then used for training and testing． The hyperparameters of the dense convolutional network is optimized before mixed dataset is classified． When the labeled data are 40% of total data volume of Automotive Hacker Dataset and a certain intrusion detection dataset, both the accuracy and recall rates are > 0.9, with error rates of < 0.1%． This method can effectively shorten detection time and reduce memory usage while detecting known and unknown attacks injected into the controller area network bus．