Dataset Card for UnlearnCanvas

This dataset card introduces "UnlearnCanvas", a high-resolution stylized image dataset for benchmarking generative modeling tasks, in particular for machine unlearning in diffusion models. Developed to address the societal concerns arising from diffusion models, such as harmful content generation, copyright disputes, and the perpetuation of stereotypes and biases, UnlearnCanvas aims at facilitating the evaluation and improvement of machine unlearning methods.

Dataset Details

UnlearnCanvas is a comprehensive, high-resolution image dataset designed to evaluate the unlearning of artistic painting styles and associated image objects. It contains images across 60 different artistic painting styles, with 400 images for each style across 20 different object categories, making it suitable for a wide range of vision generative modeling tasks beyond machine unlearning, such as style transfer, bias removal, and more.

Dataset Sources [optional]

• Repository: UnlearnCanvas GitHub
• Paper: UnlearnCanvas Paper on arXiv
• Demo: HuggingFace Benchmark

Uses

Direct Use

UnlearnCanvas is intended for direct use in:

• Evaluating machine unlearning methods for diffusion models.
• Benchmarking state-of-the-art machine unlearning techniques.
• Facilitating research in style transfer, bias removal, vision in-context learning, out-of-distribution learning, and other generative modeling tasks.

Out-of-Scope Use

• Commercial use without proper licensing or attribution may be out of scope, given the MIT license.

Dataset Structure

The dataset consists of high-resolution images across 60 different artistic painting styles, structured as ./style_name/object_name/image_idx.jpg, with a separate ./Seed_Image folder for photo-realistic images. The dataset's balanced structure and high stylistic consistency make it an ideal resource for fine-tuning and evaluating diffusion models.

Dataset Creation

Curation Rationale

The dataset was curated to address the lack of standardized and automated evaluation frameworks for machine unlearning techniques in diffusion models, facilitating the removal of undesired generative capabilities.

Source Data

Data Collection and Processing

The images were annotated (for stylization) from a set of high-resolution real-world photo-realistic images collected from the Pexels using the services provided by fotor.

Who are the source data producers?

The dataset was produced by a collaborative effort led by Yihua Zhang with contributions from their research team.

Bias, Risks, and Limitations

The dataset aims to minimize societal concerns related to diffusion models but users should be aware of the potential for misuse. Researchers are encouraged to approach the dataset with an understanding of its scope and limitations, particularly concerning the representation of styles and objects.

Recommendations

Researchers should ensure ethical use of the dataset, avoiding applications that might generate harmful content or perpetuate biases. Further studies are recommended to explore and mitigate any inherent biases within the dataset.

Citation

BibTeX:

@article{zhang2024unlearncanvas,
  title={UnlearnCanvas: A Stylized Image Dataset to Benchmark Machine Unlearning for Diffusion Models},
  author={Zhang, Yihua and Zhang, Yimeng and Yao, Yuguang and Jia, Jinghan and Liu, Jiancheng and Liu, Xiaoming and Liu, Sijia},
  journal={arXiv preprint arXiv:2402.11846},
  year={2024}
}