TY - GEN
T1 - Real-Time Privacy Guardianship in XR Metaverses via AI-Driven Behavioral Analytics
AU - Kalia, Ritika
AU - Kumar, Sudhakar
AU - Singh, Sunil K.
AU - Sharma, Samriti
AU - Chui, Kwok Tai
AU - Gupta, Brij B.
N1 - Publisher Copyright:
© 2026 IEEE.
PY - 2026
Y1 - 2026
N2 - The metaverse introduces embodied presence through avatars, creating novel privacy vulnerabilities in which behavioral patterns become primary attack surfaces for tracking, impersonation, and interaction profiling. This paper presents BehaviorShield, a real-time privacy intrusion detection system grounded in behavioral analytics applied to genuine XR motion data. The system is trained on the WhoIsAlyx VR motion dataset (9 players, 596,182 benign frames), augmented with 119,236 frames from the Project Aria Everyday Activities (AEA) dataset. Three simulated threat types, avatar tracking, behavioral impersonation, and spatial profiling, yield 993,636 annotated frames. A Random Forest classifier trained on 19 features with SMOTE resampling achieves a mean ROC-AUC of 0.9673 ± 0.0089 under player-level Leave-One-Subject-Out (LOSO) cross-validation and 0.9759 AUC on the best held-out fold. SHAP-based explainability identifies rolling motion variability as the dominant discriminator across all threat scenarios.
AB - The metaverse introduces embodied presence through avatars, creating novel privacy vulnerabilities in which behavioral patterns become primary attack surfaces for tracking, impersonation, and interaction profiling. This paper presents BehaviorShield, a real-time privacy intrusion detection system grounded in behavioral analytics applied to genuine XR motion data. The system is trained on the WhoIsAlyx VR motion dataset (9 players, 596,182 benign frames), augmented with 119,236 frames from the Project Aria Everyday Activities (AEA) dataset. Three simulated threat types, avatar tracking, behavioral impersonation, and spatial profiling, yield 993,636 annotated frames. A Random Forest classifier trained on 19 features with SMOTE resampling achieves a mean ROC-AUC of 0.9673 ± 0.0089 under player-level Leave-One-Subject-Out (LOSO) cross-validation and 0.9759 AUC on the best held-out fold. SHAP-based explainability identifies rolling motion variability as the dominant discriminator across all threat scenarios.
KW - avatar tracking
KW - behavioral analytics
KW - LOSO cross-validation
KW - metaverse security
KW - privacy protection
KW - random forest
KW - SHAP explainability
KW - VR motion data
UR - https://www.scopus.com/pages/publications/105048646019
U2 - 10.1109/ICIR70631.2026.11628368
DO - 10.1109/ICIR70631.2026.11628368
M3 - Conference contribution
AN - SCOPUS:105048646019
T3 - Proceedings - 2026 IEEE 5th International Conference on Intelligent Reality, ICIR 2026
BT - Proceedings - 2026 IEEE 5th International Conference on Intelligent Reality, ICIR 2026
T2 - 5th IEEE International Conference on Intelligent Reality, ICIR 2026
Y2 - 25 June 2026 through 26 June 2026
ER -