HarassGuard: Detecting Harassment Behaviors in Social Virtual Reality with Vision-Language Models.

Area of Science:

• Computer Science
• Human-Computer Interaction
• Virtual Reality Security

Background:

• Social Virtual Reality (VR) platforms offer immersive experiences but pose risks of online harassment.
• Current safety measures are reactive, and proactive solutions often require sensitive biometric data, raising privacy issues.

Purpose of the Study:

• To introduce HarassGuard, a novel vision-language model (VLM) system for proactive, privacy-preserving detection of physical harassment in social VR.
• To evaluate HarassGuard's effectiveness using only visual input, addressing privacy concerns associated with biometric data.

Main Methods:

• Developed an IRB-approved harassment vision dataset for social VR environments.
• Utilized prompt engineering and fine-tuned vision-language models (VLMs) to detect harassment behaviors.
• Focused on contextual information within social VR for harassment detection.

Main Results:

• HarassGuard achieved high accuracy: up to 88.09% for binary classification and 68.85% for multi-class classification.
• The system demonstrated competitive performance against established baselines like LSTM/CNN and Transformer models.
• HarassGuard required significantly fewer fine-tuning samples (200) compared to baselines (1,115).

Conclusions:

• HarassGuard offers an effective and privacy-preserving solution for detecting physical harassment in social VR.
• The system's reliance on visual input and contextual reasoning provides unique advantages over existing methods.
• This approach advances safety in immersive digital environments without compromising user privacy.