Dataset Card for Dataset Name

[Update]: we added the caption/prompt information (if there is one) in case other researchers need it. It is not used in our study though.

The dataset consists of 10K safe/unsafe images of 11 different types of unsafe content and two sources (real-world VS AI-generated).

Dataset Details

Uses

from datasets import load_dataset
dataset = load_dataset("yiting/UnsafeBench")["train"]
print(dataset[0])

{'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=1600x1435 at 0x7FB291227D90>,
'safety_label': 'Safe',
'category': 'Hate',
'source': 'Laion5B',
'text': "xxx"}

Out-of-Scope Use

This dataset is intended for research/education/responsible industrial evaluation. Any misuse is strictly prohibited.

Citation [optional]

@inproceedings{QSWBZZ24,
author = {Yiting Qu and Xinyue Shen and Yixin Wu and Michael Backes and Savvas Zannettou and Yang Zhang},
title = {{UnsafeBench: Benchmarking Image Safety Classifiers on Real-World and AI-Generated Images}},
booktitle = {{ACM SIGSAC Conference on Computer and Communications Security (CCS)}},
publisher = {ACM},
year = {2025}
}

Dataset Card Contact

yiting.qu@cispa.de