Technical Summary: G2VD

Problem Statement

Rapid advancements in AI video generation (e.g., Sora, Stable Video Diffusion, HunyuanVideo) have created significant security risks, including misinformation and identity fraud. While existing detectors perform well on in-domain data, they suffer from substantial performance degradation when applied to unseen generators. The paper identifies shortcut learning as the primary cause: detectors rely on domain-specific biases (such as generator fingerprints, compression patterns, or generation styles) rather than intrinsic forensic cues. This entanglement creates decision boundaries that fail to generalize when distribution shifts occur at the generator level. Current solutions, including forensic-oriented augmentation and multimodal large language models (MLLMs), often fail to consistently outperform specialized detectors under such shifts or produce unreliable explanations.

Methodology

The authors propose G2VD (Generalizable AI-Generated Video Detection), a framework grounded in Structural Causal Models (SCM) that treats intrinsic forensic cues as causal factors ($S$) and domain-specific bias as non-causal factors ($U$). The framework consists of two primary components:

1. Counterfactual Intervention Pipeline (CFIPipeline)

To weaken the spurious correlation between domain-specific bias and authenticity labels, G2VD constructs counterfactual samples. Ideally, a counterfactual sample $X^*_{cf}$ would pair the causal factor of a fake video ($S_f$) with the non-causal factor of a real video ($U_r$). Since $S$ and $U$ are latent and entangled, G2VD approximates this through a three-step process:

• VAE-based Reconstruction: Real videos are reconstructed using a pool of diverse Variational Autoencoders (VAEs) to introduce reconstruction-induced traces associated with intrinsic forensic cues.
• Frequency-Domain Alignment: To preserve the domain characteristics ($U_r$) while retaining reconstruction traces, the pipeline aligns the high-frequency amplitudes of the original real video with the phase and remaining spectral components of the reconstructed video. This mitigates high-frequency attenuation caused by lossy video coding.
• Pixel-Domain Alignment: Interpolation is applied to reduce residual color or contrast discrepancies, yielding the final counterfactual sample $X_{cf}$.

2. Causal Disentanglement Classifier

Building on the counterfactual intervention, the classifier learns to separate causal and non-causal representations using two domain-anchored branches and an independence constraint:

• Causal Branch: Trained to classify real videos as negative and both original fake videos and counterfactual samples (which share the fake causal factor $S_f$) as positive. This encourages the branch to capture intrinsic forensic cues ($S$) regardless of domain bias.
• Non-Causal Branch: Trained with a complementary labeling rule where counterfactual samples are labeled as real (sharing $U_r$ with original real videos). This encourages the branch to capture domain-specific bias ($U$).
• Independence Constraint: The two branches are regularized using the Hilbert-Schmidt Independence Criterion (HSIC) to minimize statistical dependence between the causal ($F_c$) and non-causal ($F_{nc}$) representations, ensuring they capture distinct factors.

During inference, the CFIPipeline and the non-causal branch are discarded; prediction relies solely on the backbone and the causal branch.

Key Contributions

1. Causal Formulation: The paper reframes AI-generated video detection by identifying the entanglement of intrinsic forensic cues and domain-specific bias as the root cause of cross-domain generalization failure.
2. G2VD Framework: It introduces a novel framework integrating reconstruction-based counterfactual intervention with domain-anchored causal disentanglement to mitigate shortcut learning.
3. Empirical Validation: Extensive experiments on four public datasets demonstrate that G2VD achieves strong cross-domain performance using only 10% of the available training data, validating the effectiveness of both the intervention pipeline and the disentanglement classifier.

Experimental Results

The framework was evaluated on four datasets: GenVidBench, GenVideo, GVD, and GVF.

• GenVidBench: In the challenging GenVidBench setting, G2VD-CLIP achieved an overall accuracy (ACC) of 91.9%, with an F1 score of 0.950 and AUC of 0.949. This represents a 25.0 percentage point improvement over the best baseline (CLIP) and significant gains over other state-of-the-art methods, particularly on difficult sources like MuseV and SVD.
• Other Datasets: G2VD variants consistently outperformed baseline methods across GenVideo, GVD, and GVF, with improvements ranging from 1.8 to 4.7 percentage points over matched backbones.
• Ablation Studies: Removing both the CFIPipeline and the causal disentanglement classifier resulted in a drastic drop in performance (e.g., down to ~61–66% ACC on GenVidBench), confirming that the intervention and disentanglement components are critical for generalization.
• Robustness: The model demonstrated robustness against Gaussian blur but showed sensitivity to aggressive JPEG compression, suggesting that compression removes forensic evidence utilized by the detector.

Significance and Claims

The paper claims that G2VD provides a practical basis for cross-domain video forensics by guiding detectors away from generator-specific shortcuts and toward intrinsic forensic cues. The authors assert that their approach effectively separates causal factors from non-causal biases, enabling reliable detection even under generator-level distribution shifts with limited training data. The work highlights that while MLLM-based methods offer explainability, specialized visual detectors with causal disentanglement mechanisms currently offer superior reliability and generalization in this domain. Future work is noted to focus on improving model robustness and calibration for real-world deployment.