Deepfake Detection Arena (DFD) Leaderboard

🎯 The Open Benchmark for Detecting AI-Generated Images

DFD-Arena is the first benchmark to address the open-source computer vision community's need for a comprehensive evaluation framework for state-of-the-art (SOTA) detection of AI-generated images.

While previous studies have focused on benchmarking the SOTA on content-specific subsets of the deepfake detection problem, e.g. human face deepfake benchmarking via DeepfakeBench, these benchmarks do not adequately account for the broad spectrum of real and generated image types seen in everyday scenarios.

🔍 Explore DFD-Arena

Learn how the framework evaluates on diverse, content-rich images with semantic balance between real and generated data:

✍️ Authorship

Both DFD-Arena and novel synthetic image datasets used for evaluation are created by BitMind.

Average Performance Metrics
{
  • "headers": [
    • "Detector",
    • "Accuracy",
    • "Precision",
    • "Recall",
    • "F1-Score",
    • "MCC"
    ],
  • "data": [
    • [
      • "NPR",
      • 0.7169,
      • 0.9193,
      • 0.5996,
      • 0.7258,
      • 0.5044
      ],
    • [
      • "UCF",
      • 0.7229,
      • 0.9436,
      • 0.592,
      • 0.7275,
      • 0.5285
      ],
    • [
      • "CAMO",
      • 0.7555,
      • 0.9442,
      • 0.647,
      • 0.7679,
      • 0.5707
      ]
    ],
  • "metadata": null
}

Dataset-specific Accuracy
Detector
CelebA-HQ
Flickr30k
ImageNet
DiffusionDB
CelebA-HQ-SDXL
CelebA-HQ-Flux
Flickr30k-SDXL
MS-COCO-Flux
CAMO
0.987
0.916
0.834
0.876
0.386
0.846
0.302
0.588