TOWARDS A SAFER METAVERSE: ANOMALY DETECTION IN AVATAR ACTIONS USING HUMAN ACTION TRANSFER LEARNING

Abstract

The Metaverse has captured global attention as a potential frontier for the internet's future. Avatar actions within this immersive digital realm mirror real-world behaviors, introducing safety concerns like cyberbullying and harmful interactions. A solution focusing on avatar action recognition and abnormal behavior detection has been proposed to address these issues. A dataset containing normal and abnormal action skeleton data was collected by extracting avatar skeleton data. However, our system does not depend solely on avatar data. To build a generalizable system, knowledge from human actions was transferred to comprehend avatars' behavior in the Metaverse. The models used proved effective in detecting the actions of different types of avatars. Furthermore, the anomaly detection model of the avatars' actions exhibited performance akin to human anomaly detection systems proposed in the literature. This affirms the feasibility of detecting avatar actions and abnormal behaviors, marking a significant stride toward ensuring safety and security within the Metaverse. The proposed solution is a crucial step in making the Metaverse a safe and secure place for all users.