Content-Independent Avatar Ownership Detection for Preventing Sockpuppet-Enabled Violations in the Social Metaverse

Social metaverse platforms, such as VRChat, allow users to create multiple accounts and control corresponding virtual avatars, offering a highly immersive social experience. However, this heightened sense of realism also amplifies certain violations, such as cyberbullying and sexual harassment, increasing their psychological impact on victims. Although these platforms impose account bans as a punitive measure, violators frequently create multiple sockpuppet accounts under false identities, such that the banning of one account does not effectively prevent continued misuse via other accounts. To address this issue, we propose AvatarJudger, a cross-account virtual avatar ownership detection method, which is more stable in social metaverse scenarios compared to traditional content-based account association approaches. Specifically, when a virtual avatar is identified as violating platform rules, AvatarJudger can detect all other avatars created by the same user. AvatarJudger implements a multi-stage avatar identity inference pipeline that utilizes knuckle movement telemetry data naturally generated during social interactions. It models personalized keystroke features of avatars in stages, considering both overall keystroke style and short-term keystroke habits, and makes a comprehensive decision. We recruit 30 volunteers and conduct a comprehensive evaluation of AvatarJudger using a dataset comprising 2,217 virtual avatars created by the participants with Microsoft HoloLens 2, Meta Quest 2, and Apple Vision Pro. Experimental results demonstrate that AvatarJudger performs effectively under various experimental settings.