VideoFakeD: Video-Based Deepfake Detection

• Igor L. O. Bastos IFood Brasil
• Anderson R. Rocha UNICAMP
• William Robson Schwartz UFMG

Resumo

This paper presents VideoFakeD, a lightweight deepfake detection pipeline that challenges the growing complexity of current state-of-the-art methods. Our approach combines a diverse set of handcrafted features, covering texture, illumination, and spectral descriptors, with a compact ViViT-based architecture capable of modeling spatial and temporal inconsistencies. Classical descriptors are used to reveal subtle manipulation artifacts, enabling our model to operate effectively with significantly fewer parameters. Despite using only 554 K to 1.6 M parameters, VideoFakeD achieves competitive results on FaceForensics++ (96.71% ACC, 97.57 AUC), and generalizes well to face antispoofing tasks (6.38 HTER, 95.62 AUC on Replay-Attack). These results highlight VideoFakeD as a highly efficient and practical solution for deepfake detection, especially suited for real-world, resource-constrained environments.

Palavras-chave: Graphics, Deepfakes, Pipelines, Lighting, Complexity theory, Faces