Game Engine Based Data Augmentation with In-game Customization and
Modeling for Malicious Behaviors Identification in Nuclear Security

Nuclear security is potentially seriously compromised by external and internal threats to nuclear facilities, such as stand-off attack and intrusion. Detection is the bottleneck of the Physical Protection System in nuclear security because delay and response will not work if detection fails. However, commonly employed detection systems rely on manual monitoring and extensive sensor deployment, which is time-consuming and costly. To this end, we introduce deep learning for vision-based automatic malicious behaviors identification, which benefits from its automaticity, generalizability, and diversity. The performance of deep learning models is highly dependent on training datasets. However, existing benchmark datasets are limited by small quantity and variety of scenes, objects, actions, and characters related to nuclear security. Besides, creating datasets from real-world filming is environmentally restrictive. Therefore, we propose to carry out data augmentation with game engine. First, we customize the buildings and interiors in the adopted game engine with a 3D map editor to simulate the environment of nuclear facilities. As objects involving malicious behaviors are not included in the resource library, we create various types of entities using modeling software with customized colors and textures. We further enhanced the conditions faced by simulated nuclear facilities by customizing environmental settings like time, weather, and season. The in-game simulated nuclear facilities allow characters with various clothes and appearances to perform multiple complex actions. All these customizations assure the diversity and quantity of datasets. Finally, we invoke the in-game camera in arbitrary distance of panoramic view, which ensures efficient and automated dataset generation. To validate the data augmentation approach, we feed the generated dataset to action recognition models for malicious behaviors identification against nuclear facilities. The experimental results demonstrate that the game engine augmented dataset allows an improved performance of malicious behaviors identification and helps get a better understanding in behavior identification.