ML Security PapersOpenFake: An Open Dataset and Platform Toward Real-World Deepfake Detection
Victor Livernoche 1,2, Akshatha Arodi 2, Andreea Musulan 2,3,4, Zachary Yang 1,2, Adam Salvail 2, Gaétan Marceau Caron 2, Jean-François Godbout 2,3, Reihaneh Rabbany 1,2
Published on arXiv
2509.09495
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
Detectors trained on OpenFake achieve near-perfect in-distribution performance and significantly outperform models trained on existing datasets on a curated in-the-wild social media test set.
OpenFake
Novel technique introduced
Deepfakes, synthetic media created using advanced AI techniques, pose a growing threat to information integrity, particularly in politically sensitive contexts. This challenge is amplified by the increasing realism of modern generative models, which our human perception study confirms are often indistinguishable from real images. Yet, existing deepfake detection benchmarks rely on outdated generators or narrowly scoped datasets (e.g., single-face imagery), limiting their utility for real-world detection. To address these gaps, we present OpenFake, a large politically grounded dataset specifically crafted for benchmarking against modern generative models with high realism, and designed to remain extensible through an innovative crowdsourced adversarial platform that continually integrates new hard examples. OpenFake comprises nearly four million total images: three million real images paired with descriptive captions and almost one million synthetic counterparts from state-of-the-art proprietary and open-source models. Detectors trained on OpenFake achieve near-perfect in-distribution performance, strong generalization to unseen generators, and high accuracy on a curated in-the-wild social media test set, significantly outperforming models trained on existing datasets. Overall, we demonstrate that with high-quality and continually updated benchmarks, automatic deepfake detection is both feasible and effective in real-world settings.
Key Contributions
- OpenFake: a ~4M-image politically grounded dataset pairing real images with synthetics from modern generators (Stable Diffusion v3.5, Flux, Ideogram, GPT Image 1), significantly more realistic than prior GAN-based benchmarks
- OpenFake Arena: a crowdsourced adversarial platform with a CLIP-based prompt-consistency gate that continually integrates hard validated examples to keep the benchmark current
- Empirical demonstration that detectors trained on OpenFake achieve near-perfect in-distribution accuracy and strong generalization to unseen generators and in-the-wild social media images, outperforming models trained on existing datasets
🛡️ Threat Analysis
The paper is fundamentally about detecting AI-generated synthetic images (deepfakes) — a canonical ML09 output integrity problem. It provides a benchmark dataset and evaluation platform explicitly for measuring deepfake detection capability against state-of-the-art generative models.