Abstract

The rapid growth of Extended Reality (XR) Internet services has raised significant security concerns, especially for immersive visual identity authentication. Deepfake-based impersonation attacks can harm user trust and data confidentiality. Traditional biometric systems, which are mainly unimodal facial recognition, are susceptible to synthetic media manipulation and adversarial spoofing. This paper proposes an immersive visual identity authentication system, coupled with a deepfake detection system that leverages multimodal feature fusion to enhance security in XR settings. The given model integrates spatial-temporal facial representations, periocular texture representations, voice spectral representations, and behavioural motion patterns via a hybrid attention-based fusion network. An evaluation was conducted on a dataset of 18,500 authentic and 17,300 deepfake XR interaction samples. Experiments show that the multimodal fusion model achieves an authentication accuracy of 98.7%, which is much higher than that of unimodal models (face-only: 92.4%; voice-only: 89.1%). The proposed deepfake detection module achieves the following precision, recall, F1-score, and false acceptance rate (FAR): 97.9%, 98.3%, 98.1%, and 1.2%, respectively, representing a 43% decrease in spoofing vulnerability compared to traditional CNN-based detectors. Additionally, real-time viability is verified through latency analysis, with an average authentication cycle processing delay of 34 ms, which is within the constraints of immersive XR services. The results suggest that multimodal feature fusion is associated with a high level of resistance to identity verification in immersive Internet ecosystems under synthetic identity manipulation. The proposed framework will contribute to a secure, scalable, and reliable authentication infrastructure for next-generation XR-enabled digital services.