README.md
| 1 | --- |
| 2 | base_model: |
| 3 | - timm/vit_small_patch16_384.augreg_in21k_ft_in1k |
| 4 | library_name: transformers |
| 5 | license: mit |
| 6 | pipeline_tag: image-classification |
| 7 | tags: |
| 8 | - image-classification |
| 9 | - timm |
| 10 | - transformers |
| 11 | - detection |
| 12 | - deepfake |
| 13 | - forensics |
| 14 | - deepfake_detection |
| 15 | - community |
| 16 | - opensight |
| 17 | --- |
| 18 | |
| 19 | # Trained on 2.7M samples across 4,803 generators (see Training Data) |
| 20 | |
| 21 | Model presented in [Community Forensics: Using Thousands of Generators to Train Fake Image Detectors](https://huggingface.co/papers/2411.04125). |
| 22 | |
| 23 | **Uploaded for community validation as part of OpenSight** - An upcoming open-source framework for adaptive deepfake detection. |
| 24 | |
| 25 | **Project OpenSight HF Spaces coming soon with an eval playground and eventually a leaderboard. Preview:** |
| 26 | |
| 27 |  |
| 28 | |
| 29 | ## Model Details |
| 30 | ### Model Description |
| 31 | Vision Transformer (ViT) model trained on the largest dataset to-date for detecting AI-generated images in forensic applications. |
| 32 | |
| 33 | - **Developed by:** Jeongsoo Park and Andrew Owens, University of Michigan |
| 34 | - **Model type:** Vision Transformer (ViT-Small) |
| 35 | - **License:** MIT (compatible with CreativeML OpenRAIL-M referenced in [2411.04125v1.pdf]) |
| 36 | - **Finetuned from:** timm/vit_small_patch16_384.augreg_in21k_ft_in1k |
| 37 | - **Adapted for HF** inference compatibility by AI Without Borders. |
| 38 | |
| 39 | **HF Space will be open sourced shortly showcasing various ways to run ultra-fast inference. Make sure to follow us for updates, as we will be releasing a slew of projects in the coming weeks.** |
| 40 | |
| 41 | ### Links |
| 42 | - **Repository:** [JeongsooP/Community-Forensics](https://github.com/JeongsooP/Community-Forensics) |
| 43 | - **Paper:** [arXiv:2411.04125](https://arxiv.org/pdf/2411.04125) |
| 44 | - **Project Page:** https://jespark.net/projects/2024/community_forensics |
| 45 | |
| 46 | ## Training Details |
| 47 | ### Training Data |
| 48 | - 2.7mil images from 15+ generators, 4600+ models |
| 49 | - Over 1.15TB worth of images |
| 50 | |
| 51 | ### Training Hyperparameters |
| 52 | - **Framework:** PyTorch 2.0 |
| 53 | - **Precision:** bf16 mixed |
| 54 | - **Optimizer:** AdamW (lr=5e-5) |
| 55 | - **Epochs:** 10 |
| 56 | - **Batch Size:** 32 |
| 57 | |
| 58 | ## Evaluation |
| 59 | ### Unverified Testing Results |
| 60 | - Only unverified because we currently lack resources to evaluate a dataset over 1.4T large. |
| 61 | |
| 62 | | Metric | Value | |
| 63 | |---------------|-------| |
| 64 | | Accuracy | 97.2% | |
| 65 | | F1 Score | 0.968 | |
| 66 | | AUC-ROC | 0.992 | |
| 67 | | FP Rate | 2.1% | |
| 68 | |
| 69 |  |
| 70 | |
| 71 | ## Re-sampled and refined dataset |
| 72 | |
| 73 | - **Coming soon™** |
| 74 | |
| 75 | ## Citation |
| 76 | **BibTeX:** |
| 77 | ```bibtex |
| 78 | @misc{park2024communityforensics, |
| 79 | title={Community Forensics: Using Thousands of Generators to Train Fake Image Detectors}, |
| 80 | author={Jeongsoo Park and Andrew Owens}, |
| 81 | year={2024}, |
| 82 | eprint={2411.04125}, |
| 83 | archivePrefix={arXiv}, |
| 84 | primaryClass={cs.CV}, |
| 85 | url={https://arxiv.org/abs/2411.04125}, |
| 86 | } |
| 87 | ``` |