HardenVR: Harassment Detection in Social Virtual Reality

Social Virtual Reality (VR) is regarded as one of the most popular VR applications since it transcends geographical barriers, allowing users to interact in simulated environments for various purposes. Despite its promising prospects, there is a growing concern about the harassment issue due to the immersive nature of social VR compared to other online social environments. Existing protections against harassment in social VR are highly limited in terms of practical effectiveness. The deficiency of studies toward understanding and preventing harassment in social VR further complicates the regulation and intervention efforts of social VR platforms in such situations. To address these challenges, we, in this paper, quantitatively investigate human interaction behaviors in social VR. More specifically, we first build a customized platform based on Mozilla Hubs, a popular social VR platform, to collect data about users' social interaction behaviors involving harassment instances. A subsequent analysis of the collected dataset SAHARA (Social interAction beHAviors in vR with hArassment) reveals that the task of online harassment detection in social VR is complicated since it depends on not only users' actions but also their spatial and temporal relationships. To accurately discern harassment, we propose a novel framework HardenVR (HA-Rassment DEtectioN framework for social VR). As a context-aware harassment detection framework, HardenVR employs a transformer-based model to capture relative poses and learn users' hand actions in 6-DOF (Degree-of-Freedom). Meanwhile, multiple mechanisms, including the extra attention mechanism, distance-aware clustering method, and the sliding window, have been introduced into the model to handle challenges of data imbalance, over-fitting, and continuous detection. The design of HardenVR aims to achieve the balance between accuracy, efficiency, and cost-effectiveness for the task of harassment detection. As a starting point, HardenVR successfully learns pose information as the context to identify harassment and the experiment results show its detection accuracy as high as 98.26%.