A Video Deepfake Detection Algorithm Based on Fusion of Physical Laws and Spatial-Temporal Features

Abstract

The emergence of high-fidelity video generation models such as Sora and Wan 2.2 has made video deepfake detection increasingly dependent on visual details. For visually degraded videos with insufficient visual details, current video deepfake detection methods suffer a substantial performance drop. In this case, incorporating physical law features into video authenticity discrimination is highly beneficial to improving the performance of deepfake detection algorithms. Based on the above analysis, this paper proposes a video deepfake detection algorithm based on fusion of physical law features and spatialtemporal features. The algorithm uses DeMamba to extract spatial-temporal features and V-JEPA to extract physical law features. Meanwhile, a degradation-aware fusion module is proposed to fuse these two features, and the fused score is utilized to identify video authenticity. Since V-JEPA works in the feature representation space rather than directly processing pixel-level details, the physical law features supported by V-JEPA can effectively complement the spatial-temporal features under visual degradation scenarios. The proposed degradation-aware fusion module adjusts the weight of each branch according to the video degradation condition. Experimental results demonstrate that our method improves the robustness of the detector under the degraded condition. From the perspective of communication security, the designed detection framework effectively strengthens deepfake detection capability under degraded transmission conditions, thereby providing reliable support for credible multimedia authentication in social-media-like degraded transmission conditions.