As the risk of cyber and safety threats to vehicle systems has increased, the anomaly detection in in-vehicle networks (IVN) has received the attention of researchers. Although, machine-learning-based anomaly detection methods have been proposed, there are limitations in detecting unknown attacks that the model has not learned because general supervised learning-based approaches depend on training dataset. To solve this problem, we propose a novel self-supervised method for IVN anomaly detection using noised pseudo normal data. The proposed method consists of two deep-learning models of the generator and the detector, which generates noised pseudo normal data and detects anomalies, respectively. Firstly, the generator is trained with only normal network traffic to generate pseudo normal traffic data. Then, the anomaly detector is trained to classify normal traffic and noised pseudo normal traffic as normal and abnormal, respectively. The experimental results demonstrate that the anomaly detection models, trained with the proposed method, not only significantly improved in the detection of unknown attacks, but also outperformed other semi-supervised learning-based methods.

Controller area network (CAN) is a de facto standard for in-vehicle network (IVN) that provides an efficient communication channel between electronic control units (ECUs). As the external connectivity of modern vehicles increase, concerns about cyber threats to in-vehicle networks are increasing. However, since global vehicle manufacturers do not release CAN specifications, it is difficult to know which CAN ID carries which vehicle parameter data and actually affects the operation of the vehicle. To solve this problem, this research proposes a method to discover CAN specification that related to major vehicle parameters using on-board diagnostics (OBD). The proposed algorithm is designed to analyze CAN bus messages by comparing with the OBD-II diagnostic responses to identify the CAN ID and their specific data field information that contains the same vehicle parameter data as the diagnostic responses. Experimental results performed on a real vehicle manufactured by a global car manufacturer show that the proposed algorithm can find CAN information related to major vehicle parameters including engine RPM, vehicle speed and throttle position and its effectiveness in vehicle security research.

Through advances in AI-based computer vision technology, the performance of modern image classification models has surpassed human perception, making them valuable in various fields. However, adversarial attacks, which involve small changes to images that are hard for humans to perceive, can cause classification models to misclassify images. Considering the availability of classification models that use neural networks, it is crucial to prevent adversarial attacks. Recent detection methods are only effective for specific attacks or cannot be applied to various models. Therefore, in this paper, we proposed an attention mechanism-based method for detecting adversarial attacks. We utilized a framework using an ensemble model, Grad-CAM and calculated the silhouette coefficient for detection. We applied this method to Resnet18, Mobilenetv2, and VGG16 classification models that were fine-tuned on the CIFAR-10 dataset. The average performance demonstrated that Mobilenetv2 achieved an F1-Score of 0.9022 and an accuracy of 0.9103, Resnet18 achieved an F1-Score of 0.9124 and an accuracy of 0.9302, and VGG16 achieved an F1-Score of 0.9185 and an accuracy of 0.9252. The results demonstrated that our method not only detects but also prevents adversarial attacks by mitigating their effects and effectively restoring labels.

With the advancement of video recording devices and network infrastructure, we use surveillance cameras to protect our valuable assets. This paper proposes a novel system for encrypting personal information within recorded surveillance videos to enhance efficiency and security. The proposed method leverages Dlib’s CNN-based facial recognition technology to identify Regions of Interest (ROIs) within the video, linking these ROIs to generate unique IDs. These IDs are then combined with a master key to create entity-specific keys, which are used to encrypt the ROIs within the video. This system supports selective decryption, effectively protecting personal information using surveillance footage. Additionally, the system overcomes the limitations of existing ROI recognition technologies by predicting unrecognized frames through post-processing. This research validates the proposed technology through experimental evaluations of execution time and post-processing techniques, ensuring comprehensive personal information protection. Guidelines for setting the thresholds used in this process are also provided. Implementing the proposed method could serve as an effective solution to security vulnerabilities that traditional approaches fail to address.