Detection of Face-Swapped Videos in Real-Time

Contemporary enterprise and consumer communications increasingly rely on live video platforms such as WhatsApp Desktop, Zoom, Microsoft Teams, and Google Meet. However, the proliferation of deep generative adversarial networks (GANs) and real-time latent auto encoders has enabled malicious actors to execute photo realistic face-swapped deepfakes at line rate, giving rise to real-time executive impersonation, identity fraud, and credential-phishing attacks. Traditional forensic detectors rely predominantly on static visual artifacts, high-frequency boundary anomalies, or pixel-level blending seams, which fail against modern generative models that synthesize coherent skin textures and seamless boundary transitions. Furthermore, heavy spatio-temporal deep neural networks incur prohibitive computational overhead that violates the stringent sub-35 ms latency constraint required for interactive video conferencing. This paper presents DeepGaze-Guard, a real-time, explainable threat intelligence framework designed specifically for cross-platform live video interactions. The system implements a dual-tier forensic architecture that couples high-throughput spatiotemporal convolutional modeling with physiological Point-of-Gaze (PoG) behavioral biometrics. Native video ingestion modules monitor hardware webcams and third-party conferencing windows via memory-mapped screen capture, incorporating Windows Per-Monitor V2 DPI-awareness .