FakeI2V-Bench: Benchmarking Image-Level Deepfake Detectors for Video
The introduction of FakeI2V-Bench marks a significant advancement in assessing deepfake video detection, particularly regarding image-level detectors' capabilities within the video sphere. This benchmark features 97,548 videos produced by cutting-edge generation models across various categories. It aims to fill the gaps in existing deepfake video detection benchmarks and the systematic evaluation of image-level detectors in video settings. The research analyzes eight video-level detectors alongside twelve notable image-level detectors. Findings reveal that the top image-level detector reaches an AUC of 80.16%, slightly surpassing some video-level counterparts, highlighting the viability of image-level detectors for detecting video deepfakes. The benchmark is detailed in a paper available on arXiv (2608.03096), emphasizing the escalating risks posed by recent advancements in video generation technology.
Key facts
- FakeI2V-Bench is a new benchmark for deepfake video detection.
- It comprises 97,548 videos.
- Videos are generated by the latest powerful generation models.
- The benchmark covers a broader range of categories.
- Eight video-level detectors and twelve image-level detectors were evaluated.
- Best-performing image-level detector achieves 80.16% AUC.
- The study is published on arXiv with ID 2608.03096.
- The benchmark aims to fill the gap in assessing image-level detectors in video domain.
Entities
Institutions
- arXiv