G2VD: Generalizable AI-Generated Video Detection via Counterfactual Intervention and Causal Disentanglement

As AI video generation models become more sophisticated, existing detectors often struggle to generalize to unseen generators. This is largely due to shortcut learning, where models rely on superficial domain-specific biases (such as generator-specific fingerprints or compression styles) rather than the intrinsic forensic cues that define AI-generated content. This paper investigates how to improve cross-domain generalization by causally separating these forensic cues from domain-specific noise.

The authors propose G2VD, a framework based on structural causal modeling. It consists of two primary innovations:

1. Counterfactual Intervention Pipeline (CFIPipeline): This module constructs counterfactual samples by combining VAE-based reconstruction with frequency-domain and pixel-domain alignment. By pairing fake-video forensic cues with real-video domain characteristics, the pipeline weakens the spurious correlation between domain-specific bias and authenticity labels.
2. Causal Disentanglement Classifier: This classifier uses two domain-anchored branches—one for causal (forensic) features and one for non-causal (domain-specific) features. It employs the Hilbert-Schmidt Independence Criterion (HSIC) as a constraint to ensure the two branches capture distinct, non-overlapping information, effectively forcing the model to focus on intrinsic forensic cues.

G2VD demonstrates robust cross-domain performance across four public datasets, consistently outperforming state-of-the-art baselines. On the challenging GenVidBench benchmark, G2VD achieved over 90% accuracy, providing significant improvements in F1 and AUC scores while using only 10% of the available training data. Ablation studies confirm that both the counterfactual intervention and the causal disentanglement components are essential for achieving these gains, with the intervention providing the most substantial boost in generalization.