Paper summary: Can We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? A Systematic Evaluation of Detectors, Generators and Social Dissemination

Problem

Video generators can now fabricate realistic footage of wars, disasters and emergencies, a serious misinformation risk. Existing benchmarks say little about how detectors behave in exactly these settings.

Method

RA-Bench uses real videos as anchors: 1,830 real videos across 10 social-risk categories, plus 16,056 generated clips from four open-source and five closed-source generators. The evaluation has three parts:

1. detector generalisation: seven traditional detectors, ten zero-shot multimodal models in three review settings, and two MLLMs fine-tuned for AI-video detection
2. how detectability changes with generation quality, conditioning and seeds
3. human authenticity judgements, and detector reliability during social dissemination (re-encoding and re-sharing)

Key results

From the abstract: none of the three detector families generalises consistently. Generation properties affect the families differently, while source-level detection patterns stay stable across seeds.

Limitations

There’s no dedicated limitations section in the extracted text. Generators change quickly, so any snapshot benchmark ages.