AI-Generated Video Detection

End-to-end pipeline for detecting AI-generated videos with 99%+ accuracy across Sora, Runway AI, Stable Video Diffusion, and NeRF-based systems. Published at CVPR 2024.

Overview

While deepfake detection research has focused heavily on images, AI-generated videos present a fundamentally different forensic challenge. Generators like Sora, Runway AI, and Stable Video Diffusion introduce traces that are qualitatively different from image generators — and existing image detectors fail to catch them.

Approach

We engineered an end-to-end detection pipeline that exploits temporal inconsistencies and generator-specific forensic artifacts unique to video generation. The system uses CNNs with temporal forensic analysis to identify AI-generated content across multiple generative architectures.

Dataset

We constructed and released a large-scale benchmark dataset of 8M+ frames spanning:

  • Transformer-based generators (Sora)
  • Diffusion-based generators (Runway AI, Stable Video Diffusion)
  • NeRF-based systems

Hosted on Hugging Face with 400+ community downloads and adopted by external research groups for benchmarking.

Results

  • 99%+ accuracy across 4 major generator families
  • Robust detection after H.264 re-compression
  • Few-shot learning enables detection of new generators with minimal examples

Impact

  • Published at CVPR Workshops 2024
  • Dataset adopted by external research groups
  • 400+ Hugging Face downloads