G2VD targets shortcut learning in AI-generated video detectors, which often perform poorly on videos from unseen generators. Its counterfactual intervention pipeline uses VAE reconstruction followed by frequency- and pixel-domain alignment to weaken correlations between generator-specific biases and authenticity labels. A two-branch causal disentanglement classifier, regularized with the Hilbert-Schmidt Independence Criterion, separates intrinsic forensic cues from domain-specific information. The authors report consistent cross-domain gains across four public datasets and more than 90% overall accuracy on GenVidBench using only 10% of the available training data.
No heat snapshots are available in the last 24 hours.