ACID: A Comprehensive Dataset for AI-Created Image Detection
Keywords: Computer vision, Generative Model, AI Ethics
TL;DR: We present a benchmark for AI generated image detection.
Abstract: Generative models have demonstrated remarkable capabilities in generating photorealistic images under proper conditional guidance. Such advancements raise concerns about potential negative social impacts, such as the proliferation of fake news. In response, numerous methods have been developed to differentiate fake from real. Yet, their accuracy and reliability still need to be improved, especially when facing state-of-the-art generative models such as large diffusion models. Infrastructure-wise, the existing testing datasets are sub-optimal in terms of research dimensions and product utility due to their limited data volume and insufficient domain diversity.
In this work, we introduce a comprehensive new dataset, namely ACID, which consists of 13M samples sourced from over 50 different generative models versus real-world scenarios. The AI-generated images in this collection are sampled based on fine-grained text prompts and span multiple resolutions. For the real-world samples, we broadly searched public data sources and carefully filtered text-image pairs based on visual and caption quality.
Using ACID, we present ACIDNet, an effective framework for detecting AI-generated images. ACIDNet leverages texture features from a Single Simple Patch (SSP) branch and semantic features from a ResNeXt50 branch, and achieves overall cross-benchmark accuracy of $86.77\%$, significantly outperforming previous methods such as SSP and CNNSpot by over $10\%$. Both our model and dataset will be open-released to the public.
Primary Area: datasets and benchmarks
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1941
Loading