ML Security PapersPhysics-Driven Spatiotemporal Modeling for AI-Generated Video Detection
Shuhai Zhang 1,2, ZiHao Lian 1, Jiahao Yang 1, Daiyuan Li 1, Guoxuan Pang 3, Feng Liu 4, Bo Han 5, Shutao Li 6, Mingkui Tan 1
1 South China University of Technology
3 University of Science and Technology of China
Published on arXiv
2510.08073
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
NSG-VD outperforms state-of-the-art baselines by 16.00% in Recall and 10.75% in F1-Score on AI-generated video detection benchmarks
NSG-VD
Novel technique introduced
AI-generated videos have achieved near-perfect visual realism (e.g., Sora), urgently necessitating reliable detection mechanisms. However, detecting such videos faces significant challenges in modeling high-dimensional spatiotemporal dynamics and identifying subtle anomalies that violate physical laws. In this paper, we propose a physics-driven AI-generated video detection paradigm based on probability flow conservation principles. Specifically, we propose a statistic called Normalized Spatiotemporal Gradient (NSG), which quantifies the ratio of spatial probability gradients to temporal density changes, explicitly capturing deviations from natural video dynamics. Leveraging pre-trained diffusion models, we develop an NSG estimator through spatial gradients approximation and motion-aware temporal modeling without complex motion decomposition while preserving physical constraints. Building on this, we propose an NSG-based video detection method (NSG-VD) that computes the Maximum Mean Discrepancy (MMD) between NSG features of the test and real videos as a detection metric. Last, we derive an upper bound of NSG feature distances between real and generated videos, proving that generated videos exhibit amplified discrepancies due to distributional shifts. Extensive experiments confirm that NSG-VD outperforms state-of-the-art baselines by 16.00% in Recall and 10.75% in F1-Score, validating the superior performance of NSG-VD. The source code is available at https://github.com/ZSHsh98/NSG-VD.
Key Contributions
- Normalized Spatiotemporal Gradient (NSG) statistic grounded in probability flow conservation principles that quantifies deviations from natural video spatiotemporal dynamics
- NSG estimator leveraging pre-trained diffusion model score functions for spatial gradient approximation and brightness constancy for temporal modeling, avoiding explicit optical flow computation
- NSG-VD detection method using Maximum Mean Discrepancy between NSG features of test and real videos, with theoretical upper bound proving generated videos exhibit amplified discrepancies
🛡️ Threat Analysis
Primary contribution is a novel AI-generated content detection method (NSG-VD) for synthetic videos. The paper introduces a new forensic technique grounded in probability flow conservation physics, leveraging diffusion model score functions to detect deepfake/synthetic videos — directly addressing output integrity and content authenticity.