karthik-2905/OpenFake

---
dataset_info:
features:
- name: image
dtype: image
- name: prompt
dtype: string
- name: label
dtype: string
- name: model
dtype: string
splits:
- name: train
num_bytes: 1051540257907.984
num_examples: 1870684
- name: test
num_bytes: 33418712589.0
num_examples: 59658
download_size: 1083933904266
dataset_size: 1084958970496.984
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
license: cc-by-sa-4.0
task_categories:
- image-classification
language:
- en
size_categories:
- 100K<n<1M
---

# Dataset Card for OpenFake

## Dataset Details

### Dataset Description

OpenFake is a dataset designed for evaluating deepfake detection and misinformation mitigation in the context of politically relevant media. It includes high-resolution real and synthetic images generated from prompts with political relevance, including faces of public figures, events (e.g., disasters, protests), and multimodal meme-style images with text overlays. Each image includes structured metadata with its prompt, source model (for synthetic), and human-annotated or pipeline-assigned labels.

* **Authors:** Victor Livernoche; Akshatha Arodi; Andreea Musulan; Zachary Yang; Adam Salvail; Gaétan Marceau Caron; Jean-François Godbout; Reihaneh Rabbany
* **Curated by:** Victor Livernoche; Akshatha Arodi; Jie Zang
* **Funded by:** CIFAR AI Chairs Program; Centre for the Study of Democratic Citizenship (CSDC); IVADO; Canada First Research Excellence Fund; Mila (financial support and computational resources)
* **Language(s) (prompts):** English
* **License:** CC-BY-SA-4.0. Note: subsets produced with proprietary generators are released under non-commercial terms due to “non-compete” clauses; see paper for details.

### Dataset Sources

- **Repository:** [https://huggingface.co/datasets/ComplexDataLab/OpenFake](https://huggingface.co/datasets/ComplexDataLab/OpenFake)
- **Arena (crowdsourced adversarial platform):** [https://huggingface.co/spaces/CDL-AMLRT/OpenFakeArena](https://huggingface.co/spaces/CDL-AMLRT/OpenFakeArena)

## Uses

### Direct Use

* Benchmarking binary classifiers for real vs. synthetic image detection
* Evaluating robustness across models and content types (faces, events, memes)
* Training adversarially robust detectors via community submissions (OpenFake Arena)

### Out-of-Scope Use

* Training generative models directly on the dataset without consent
* Any use of personal imagery that violates platform rules or privacy

## Dataset Structure

* `image`: image (real or synthetic)
* `label`: `real` or `fake`
* `model`: the model that generated the synthetic image
* `prompt`: prompt used to generate the synthetic image or caption for a real image

Train/test split is balanced by label and curated for visual and topical diversity. No image overlaps between splits.

**Unused metadata:** `unused_metadata.csv` contains URLs and prompts for images not included in the train/test splits.

## Models Covered

Synthetic images were generated from a diverse set of state-of-the-art generators, including:

- Stable Diffusion **1.5**, **2.1**, **XL**, **3.5**
- Flux **1.0-dev**, **1.1-Pro**, **1.0-Schnell**
- Midjourney **v6**, **v7**
- **DALL·E 3**, **Imagen 3**, **Imagen 4**
- **GPT Image 1**, **Ideogram 3.0**, **Grok-2**, **HiDream-I1**, **Recraft v3**, **Chroma**
- Plus 10 community LoRA/finetuned variants of SD 1.5/XL and Flux-dev

All images are produced at ~1 MP with varied aspect ratios reflecting common social-media formats.

## Dataset Creation

### Curation Rationale

The goal is to fill a gap in deepfake detection datasets by covering high-quality, politically sensitive synthetic imagery and going beyond face-only benchmarks to include events and hybrid image-text memes. The dataset pairs ~3M politically themed real images (filtered from LAION-400M using Qwen2.5-VL) with ~963k synthetic counterparts, and is complemented by the OpenFake Arena for continual hard negative generation.

### Source Data

**Real images.** Selected from LAION-400M and filtered with Qwen2.5-VL to retain faces and politically salient or newsworthy events. Detailed captions are produced to drive T2I generation and Arena prompts.

**Synthetic images.** Generated using the model list above from a shared prompt bank. Open-source models follow documented generation settings for reproducibility.

#### Who are the source data producers?

* Real: news outlets, political users, and public social-media posts
* Synthetic: produced by researchers and community contributors from prompts; Arena submissions are gated by CLIP for prompt relevance and logged with metadata

#### Personal and Sensitive Information

Source data was filtered to reduce personal or sensitive content; see the paper’s ethics and licensing notes.

## Bias, Risks, and Limitations

There may be overrepresentation of Western political events due to source distribution. Synthetic examples inherit generator biases. Not all labels are exhaustively human-verified. Adversarial use is a risk, mitigated by licensing and the dataset’s focus on detection.

### Recommendations

Use caution when interpreting political narratives in images. Do not use for content generation or facial identity research without additional review.

## Citation

**BibTeX:**
```bibtex
@misc{livernoche2025openfakeopendatasetplatform,
title={OpenFake: An Open Dataset and Platform Toward Large-Scale Deepfake Detection},
author={Victor Livernoche and Akshatha Arodi and Andreea Musulan and Zachary Yang and Adam Salvail and Gaétan Marceau Caron and Jean-François Godbout and Reihaneh Rabbany},
year={2025},
eprint={2509.09495},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2509.09495},
}
```

**APA:**

Livernoche, V., Arodi, A., Musulan, A., Yang, Z., Salvail, A., Marceau Caron, G., Godbout, J.-F., & Rabbany, R. (2025). OpenFake: An open dataset and platform toward large-scale deepfake detection. arXiv. https://arxiv.org/abs/2509.09495

## More Information

For questions, errors, or contributions, visit the GitHub or HF repository.

## Dataset Card Authors

Victor Livernoche

## Dataset Card Contact

[e-mail]