RA-Bench: New Benchmark Evaluates AI-Generated Video Detection in Crisis Scenarios
A new evaluation standard called RA-Bench has been launched to assess the identification of AI-generated videos that portray real-world crises, including wars, disasters, and public emergencies. This benchmark features a total of 17,886 videos, which consist of 1,830 authentic video anchors categorized into 10 social-risk groups, alongside 16,056 generated clips sourced from four open-source and five closed-source generators. The assessment focuses on three key areas: the generalization of detectors across seven traditional models and ten zero-shot multimodal models, generation conditions, and social dissemination. This research seeks to fill gaps in understanding detector and generator performance in crisis scenarios, including detectability variations, human perceptions, and detector reliability during dissemination. The findings are available on arXiv under identifier 2608.14391.
Key facts
- RA-Bench is a new benchmark for AI-generated video detection.
- It uses real videos as anchors.
- The benchmark contains 17,886 videos.
- It includes 1,830 real-video anchors across 10 social-risk categories.
- It includes 16,056 generated clips from four open-source and five closed-source generators.
- Evaluation covers detector generalization, generation conditions, and social dissemination.
- Seven traditional detectors and ten zero-shot multimodal models are assessed.
- The study is published on arXiv with identifier 2608.14391.
Entities
Institutions
- arXiv