Large-scale Chinese data evaluation benchmark for face video forgery detection

The world's first Chinese facial video anti-counterfeiting dataset containing 434727 samples was released as a result. The experiment showed that the dataset had high difficulty in identification, and the accuracy of visual and audio-visual combination was controlled below 85% and 70% in 16 evaluations including SOTA and mainstream anti-counterfeiting models, respectively. The constructed evaluation benchmark covers both visual and auditory modal scenes, and shows an average fluctuation of 19% in model accuracy performance in cross domain generalization testing 6%, significantly revealing the application limitations of existing algorithms.

Conclusion: The constructed Chinese anti-counterfeiting evaluation benchmark effectively fills the gap in the field. Through systematic experiments, the correlation mechanism between dataset characteristics and algorithm performance is elucidated, and key development directions such as enhancing model robustness and cross modal generalization ability are proposed. This provides data support and practical guidance for quantitative evaluation in Chinese scenarios and the actual deployment of facial video anti-counterfeiting technology.