EVDO: An Enhanced Framework for Deepfake Detection in Videos Through Optical Flow and Temporal-Spatial Analysis

MM Melouk and MH Shao and A Basit and C Abouzahir and RL Zhou and M Shafique, IEEE ACCESS, 13, 169367-169380 (2025).

DOI: 10.1109/ACCESS.2025.3614455

As generative AI advances, the realism of synthetic media has sparked serious concerns in security, privacy, and misinformation. This issue is particularly concerning with the rise of deepfake technologies that manipulate facial imagery, undermining media authenticity. While existing research has largely focused on image-based deepfake detection with specialized feature extractors, detecting deepfakes in video, especially with broad generalization across varied manipulation techniques, remains a substantial challenge. This paper introduces EVDO, a novel framework designed to detect deepfake videos by integrating temporal and spatial analysis for enhanced detection accuracy. Our approach leverages optical flow techniques to capture subtle temporal manipulation artifacts between video frames overlooked in spatial analysis. Using a FlowFormer++ model for temporal analysis, frame pairs are sampled to produce cost volumes that highlight essential motion regions and capture manipulation-specific artifacts through optical flow images which encode temporal dependencies. A flow-finetuned detector then extracts flow-level features indicative of deepfake manipulation. Complementing this, Xception-based spatial detectors analyze each frame individually, generating high-dimensional embeddings that capture frame- specific anomalies. Fusing these temporal and spatial embeddings enables comprehensive binary classification of deepfakes. Validated on the FaceForensics++ benchmark, EVDO significantly improves generalization. Our proposed temporal path contributes correct classifications to around 12% of datapoints. EVDO achieves a video AUC of 99.13% (99.43% of end- to-end SOTA methods) while enabling forensically verifiable manipulation detection.

Return to Publications page