DBINDS: detection based on initial noise difference sequence from diffusion model inversion for AI-generated videos
TL;DR
DBINDS, a diffusion-model-inversion-based detection framework that extends the analysis from the pixel domain to a diffusion-inversion-derived latent-noise space, is proposed and a composite of spatiotemporal correlation and spatiotemporal texture features is identified as the Best Dual Combination.
Abstract
AI-generated video has advanced rapidly, posing serious challenges to content security and forensic analysis. Existing detection methods primarily rely on pixel-level visual features and often show limited generalization to unseen generators. We propose DBINDS, a diffusion-model-inversion-based detection framework that extends the analysis from the pixel domain to a diffusion-inversion-derived latent-noise space. DBINDS applies a fixed diffusion-inversion backbone as a detector-side analysis operator, obtains surrogate initial-noise descriptors for video frames, and constructs the Initial Noise Difference Sequence (INDS) to characterize inter-frame variations. Based on multidimensional and multiscale INDS analysis, we identify a composite of spatiotemporal correlation and spatiotemporal texture features as the Best Dual Combination. Using Bayesian hyperparameter optimization and a LightGBM classifier, we validate DBINDS on GenVidBench under a one-to-many protocol, where the model is trained on one generated source and one real source and tested on unseen generators and an unseen real-video source. The Best Dual Combination achieves 78.08% overall accuracy on the unified open-set test set. Additional ablation, reduced-data, robustness, and source-controlled cross-validation experiments further support the effectiveness and transferable detection potential of INDS as an exploratory latent cue for AI-generated-video detection.