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

Sep 21, 2026

Uploaded by Rowan

The rapid growth of AI-generated video increases security exposure and demands reliable detectors with strong cross-domain generalization. Existing approaches often perform well within known generators but degrade on unseen ones due to shortcut learning and domain-specific spurious cues. G2VD introduces a counterfactual intervention pipeline using VAEs plus frequency- and pixel-domain alignment to shift focus toward generator-intrinsic forgery traces. It also adds a causal disentanglement classifier with HSIC-based independence to separate task-relevant evidence from domain bias, achieving over 90% accuracy on GenVidBench and near-0.95 AUC using only 10% training data.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Abstract

I. Introduction

Security risks from modern AI video generation

Limits of existing detectors under generator shifts

Shortcut learning and spurious correlations

G2VD: causal perspective with counterfactual intervention

G2VD: Generalizable AI-Generated Video Detection via Counterfactual Intervention and Causal Disentanglement Meng Du , Hongchang Chen, Ran Li, Junjie Zhang, Qi Ouyang, and Shuxin Liu* arXiv :2607 .04607v 1 [ cs .CV] 6 Jul 2026 Abstract—The rapid advancement of AI-generated videos poses increasing security risks and calls for robust detectors with strong cross-domain generalization. Although existing methods achieve promising results under in-domain evaluation, their performance often degrades substantially when tested on unseen generators. A key reason is shortcut learning, where detectors rely on domain-specific spurious cues, such as generator-dependent fingerprints and generation styles, instead of intrinsic forgery traces. To address this issue, we propose G2VD, a Generalizable AI-Generated Video Detection framework based on counterfactual intervention and causal disentanglement. First, G2VD introduces a counterfactual intervention pipeline (CFIPipeline) that generates controlled counterfactual samples via variational autoencoders (VAEs), followed by frequency-domain and pixeldomain alignment, thereby encouraging the detector to focus on generator-intrinsic cues. Building on this intervention process, we further design a causal disentanglement classifier consisting of two domain-anchored branches with distinct classification objectives, combined with an HSIC-based independence constraint to encourage the separation of task-relevant cues from domainspecific bias. Across four public datasets, G2VD shows strong average cross-domain performance and consistent gains over matched backbones. On the challenging GenVidBench crossdomain setting, it exceeds 90% accuracy and reaches an AUC close to 0.95. Notably, this performance is obtained using only 10% of the original training data. The code is available at [https://github.com/dumeng98/G2VD](https://github.com/dumeng98/G2VD). Index Terms—AI-generated video detection, domain generalization, causal representation learning, counterfactual intervention, feature disentanglement. I. INTRODUCTION MODERN AI video generation models, such as Sora [1], Stable Video Diffusion [2], CogVideoX [3], HunyuanVideo [4], and Wan [5], have made it increasingly accessible to synthesize highly realistic videos from simple text or image inputs. While these technologies offer substantial creative potential, they also introduce serious security risks, including misinformation dissemination, identity fraud, and public opinion manipulation. As a result, developing reliable AI-generated video detectors with strong cross-domain generalization has become increasingly urgent. Existing detection methods have achieved promising performance under in-domain evaluation, but often degrade substantially on videos produced by unseen generators [6], [7] . Early deepfake detectors mainly focus on facial forgeries by exploiting spatial artifacts [8], [9], blending boundaries [10], reconstruction discrepancies [11], and temporal inconsistencies [12],[13] . With the rise of general-purpose video genera Meng Du, Hongchang Chen, Ran Li, Junjie Zhang, Qi Ouyang, and Shuxin Liu are with Information Engineering University, Zhengzhou 450003, China (e-mail: [dm](dm csy@alumni.nudt.edu.cn)[ ](dm csy@alumni.nudt.edu.cn)[csy@alumni.nudt.edu.cn](dm csy@alumni.nudt.edu.cn); [liushuxin11@126.com](liushuxin11@126.com)). *Corresponding author: Shuxin Liu. Fig. 1. Motivation of G2VD. Conventional training entangles causal forgery cues with non-causal domain biases, yielding a boundary that fits seen generators but fails on unseen ones. Counterfactual intervention approximately preserves real-video domain factors while introducing generator-like reconstruction cues, weakening their spurious correlation with labels. Causal disentanglement further encourages the separation of causal and non-causal representations, producing a boundary that transfers to unseen generators. tors, recent efforts extend detection to arbitrary

FAQs

What causes detectors to fail when evaluated on unseen video generators?

Performance drops because shortcut learning leads detectors to rely on domain-specific spurious cues (e.g., generator fingerprints and styles) instead of intrinsic forgery traces.

How does G2VD use counterfactual intervention to improve generalization?

G2VD introduces a counterfactual intervention pipeline that generates controlled counterfactual samples with VAEs and performs frequency-domain and pixel-domain alignment to weaken spurious correlations with labels.

What is the role of the causal disentanglement classifier in G2VD?

It uses two domain-anchored branches with distinct objectives and an HSIC-based independence constraint to encourage separation of task-relevant causal cues from domain-specific bias.

Sep 21, 2026

Uploaded by Rowan

G2VD: Generalizable AI-Generated Video Detection via Counterfactual Intervention and Causal Disentanglement_Page 1

Open in WPS to view all 10 pages