ReConFuse: Reconstruction-Error Guided Semantic Fusion for AI-Generated Video Detection
Researchers have developed a new method called ReConFuse to detect AI-generated videos by analyzing reconstruction errors. This technique leverages a pre-trained Weighted Finite Variational Autoencoder (WF-VAE) to identify distinct patterns in frame-wise reconstruction errors between real and AI-generated videos. The framework then fuses these error cues with semantic features and uses a Mamba-based module to model temporal dynamics for video-level classification, demonstrating effectiveness and generalization across various generative models. AI
IMPACT This method could improve the detection of AI-generated videos, aiding in combating misinformation and ensuring media authenticity.