Self-Supervised AI-Generated Image Detection: A Camera Metadata Perspective

Official implementation of the paper "Self-Supervised AI-Generated Image Detection: A Camera Metadata Perspective".

⭐️Our series work: BLADES (ICCV'25 Highlight)

🆕 News

Our paper has been accepted for publication in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 🎉🎉🎉

📜 Abstract

The proliferation of AI-generated imagery poses escalating challenges for multimedia forensics, yet many existing detectors depend on assumptions about the internals of specific generative models, limiting their cross-model applicability. We introduce a self-supervised approach for detecting AI-generated images that leverages camera metadata---specifically exchangeable image file format (EXIF) tags---to learn features intrinsic to digital photography. Our pretext task trains a feature extractor solely on camera-captured photographs by classifying categorical EXIF tags (\eg, camera model and scene type) and pairwise-ranking ordinal and continuous EXIF tags (\eg, focal length and aperture value). Using these EXIF-induced features, we first perform one-class detection by modeling the distribution of photographic images with a Gaussian mixture model and flagging low-likelihood samples as AI-generated. We then extend to binary detection that treats the learned extractor as a strong regularizer for a classifier of the same architecture, operating on high-frequency residuals from spatially scrambled patches. Extensive experiments across various generative models demonstrate that our EXIF-induced detectors substantially advance the state of the art, delivering strong generalization to in-the-wild samples and robustness to common benign image perturbations.

---

Key Contributions:

• A self-supervised pretext task that leverages EXIF tags to learn camera-intrinsic features from photographs only.
• A feature extractor that operates on high-frequency residuals of scrambled patches to suppress semantics and accentuate imaging regularities.
• A one-class detector that models photographic features with a GMM and detects anomalies without seeing AI-generated images during training.
• A binary detector that uses the pretext extractor as a strong regularizer, improving generalization and robustness across generators and post-processing.

---

🛠️ Environment Configuration

This project recommends using uv for fast package management:

uv sync

📂 Dataset Description & Downloads

Parameter	Description	Resource Link
-exif_image_path	Images with EXIF metadata required for backbone training	Download
-test_image_path	Test sets	Download
-oc_realonly_image_path	One-class training set (contains only ImageNet/LSUN real images)	Download
-bc_trainset_path	Training dataset for binary classification	Download

🚀 Quick Start

1. Training the Backbone

torchrun --nproc_per_node=8 backbone_train.py

2. One-Class Detection Pipeline (Backbone Evaluation)

python oc_main.py

3. Binary Classification Model Training

torchrun --nproc_per_node=8 bc_train.py

4. Binary Classification Model Evaluation

python bc_eval.py

📦 Pre-trained Weights

We provide pre-trained model checkpoints for quick reproduction: backbone checkpoint and BC classifier checkpoint.

✍️ Citation

If you find our work or code useful for your research, please cite:

@article{zhong2026self,
  title={Self-Supervised AI-Generated Image Detection: A Camera Metadata Perspective},
  author={Zhong, Nan and Zou, Mian and Xu, Yiran and Qian, Zhenxing and Zhang, Xinpeng and Wu, Baoyuan and Ma, Kede},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2026}
}

@inproceedings{zou2025bi,
  title={Bi-Level Optimization for Self-Supervised AI-Generated Face Detection},
  author={Zou, Mian and Zhong, Nan and Yu, Baosheng and Zhan, Yibing and Ma, Kede},
  year={2025},
  booktitle={International Conference on Computer Vision},
  pages={18959--18968}
}