Micro-Expression-Aware Avatar Fingerprinting via Inter-Frame Feature Differencing
Researchers have developed a new method for avatar fingerprinting that verifies who is controlling a synthetic talking-head video. This system operates directly on raw video frames without preprocessing, utilizing micro-expression awareness and inter-frame feature differencing. By subtracting consecutive feature maps, the model preserves driver-specific motion dynamics while minimizing the impact of stable appearance features. Experiments on the NVFAIR dataset showed the system achieved an AUC of 0.877, outperforming landmark-based methods on several cross-generator pairs. AI
IMPACT Enhances security for synthetic media by enabling verification of avatar control, potentially impacting content moderation and digital identity.