OpenFake: An Open Dataset and Platform Toward Large-Scale Deepfake Detection (arxiv.org)

Researchers released OpenFake, a large-scale, politically-focused dataset and accompanying crowdsourced platform designed to push deepfake detection toward real-world, modern threats. The dataset pairs three million real images with descriptive captions and uses those captions to generate 963k high-quality synthetic images produced by a mix of proprietary and open-source generative models. Motivated by social-media analysis and a human perception study showing that recent proprietary generators produce images increasingly indistinguishable from real photos, the project explicitly targets multiple dissemination modalities (not just single-face images) to better reflect how synthetic content spreads online. Technically, OpenFake fills gaps in prior benchmarks by offering scale, caption-conditioned synthetic creation, and a continuously updated adversarial pipeline: contributors are incentivized to submit hard-to-detect fakes, keeping the benchmark aligned with evolving generative capabilities. For the AI/ML community this means richer training data and more realistic evaluation — exposing detectors to high-fidelity, multimodal attacks and enabling stress-testing against proprietary models — while also raising dual-use and privacy considerations that call for responsible access and governance. Overall, OpenFake aims to make detection methods more robust and adaptive to the current generation of generative models.