AIAny
Icon for item

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

Systematically evaluates AI-generated video detectors and generators for real-world crisis scenarios using RA-Bench (17,886 clips: 1,830 real anchors, 16,056 generated). Shows detector families fail to generalize across generation conditions, and that human-misleading videos and social dissemination further degrade detection.

Introduction

Why this matters

AI video generators can fabricate realistic depictions of wars, disasters and other crises that spread rapidly on social platforms. Detecting such videos is not just an artifact-matching problem: detectors must generalize across diverse generators, conditioning methods, and the social processes that amplify misinformation. This paper provides a focused empirical lens on those real-world challenges by building RA-Bench and evaluating detectors, generators, and dissemination together.

Key Findings
  • RA-Bench and evaluation design: introduces RA-Bench with 17,886 videos (1,830 real-video anchors across 10 social-risk categories and 16,056 generated clips) produced by four open-source and five closed-source generators. The benchmark organizes analysis along three dimensions: detector generalization, generation-condition sensitivity, and social dissemination effects.

  • Detector generalization is weak: evaluates seven traditional detectors, ten zero-shot multimodal models under three review settings, and two MLLMs fine-tuned for generated-video detection. None of these families generalizes consistently across RA-Bench instances; performance varies substantially by generator and scenario.

  • Generation properties matter but patterns are stable by source: detectability depends on generation quality, conditioning information, and sampling seeds; however, source-level detection patterns remain relatively stable across seeds, implying generator-level biases.

  • Human judgments and dissemination interact with detectors: videos that frequently mislead humans are also harder for current detectors, and detection reliability degrades further after simulated social dissemination, highlighting a gap between lab benchmarks and field risk.

  • Practical implication: improving robustness requires evidence-based, semantics-aware verification (beyond low-level artifacts), broader generator coverage in training/benchmarks, and evaluation under realistic social pipelines.

Who it's for and tradeoffs

Great fit if you are a researcher or practitioner building or evaluating AI-generated video detectors, a dataset/benchmark developer aiming to capture real-world risk categories, or a policymaker assessing platform-level defenses. The study provides concrete failure modes (generator-dependent gaps, dissemination-induced degradation) and a sizable public benchmark to reproduce analyses.

Look elsewhere if you need a narrow artifact detector tuned to a specific generator or if your threat model excludes social dissemination: RA-Bench emphasizes cross-generator realism and social-context evaluation, so its findings may underrepresent scenarios limited to a single, well-known generator. The benchmark is comprehensive across many generators and categories but remains constrained to the included models, categories, and the authors' dissemination simulations.

Information

  • Websitearxiv.org
  • AuthorsShuo Liang, Yixing Ma, Pengfei Zhou, Xingyan Chen, Zihan Mei, Manting Li, Feihan Chen, Zhiwen Wang, Bin Xu, Haotian Zhang
  • Published date2026/08/14

Categories

More Items

Externalizes persistent scene state into a camera-indexed world bank and designs a long-horizon teacher whose sparse-attention supervision is distilled into a three-step student, enabling responsive, low-latency interactive long-horizon video generation with bounded denoiser context.

Predicts future video frames conditioned on an observed frame, a language instruction, and a sequence of end-effector poses and gripper states for robot manipulation. Uses per-arm SE(3) geometric encoding (PRoPE-style), a lightweight depth branch, SAM3 masks with a frozen V-JEPA teacher, and distribution-matching distillation for efficient, consistent action-conditioned rollouts.

Uses video generation only as a training signal to co-train a pretrained video expert and a lightweight action expert, then discards the video branch at inference to produce a low-latency end-to-end driving planner; enhanced with RL for compositional driving rewards.