Beyond Face Swapping: A Diffusion-Based Digital Human Benchmark for Multimodal Deepfake Detection

Recent advances in deepfake technologies, particularly digital human generation, pose escalating threats due to their realism and covert multimodal control. Existing detection methods struggle with these new challenges. To address this gap, we first present DigiFakeAV<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>, a large-scale, multimodal forgery dataset featuring 60,000 videos generated by five state-of-theart digital human models, ensuring diversity in nationality, skin tone, gender, and scenario. Existing detectors exhibit marked performance drops on DigiFakeAV, underlining its difficulty. To counter this, we propose DigiShield, a robust baseline leveraging spatiotemporal and cross-modal fusion of visual and audio features. DigiShield attains state-of-the-art results on DigiFakeAV and generalizes well to other datasets.