A Standard-Driven and Fine-Grained Dataset for Content Safety Detection of AI-Generated Images

Specifically, using standard terms as the sole semantic source, abstract compliance requirements are transformed into actionable sets of risk themes through four stages: term mapping, topic expansion, prompt word synthesis, and image generation; Using multiple cultural image models such as Qwen Image, Hunyuan Image, Kolors, and Z-Image to synthesize images, in order to expand the diversity of samples in style, composition, and imaging habits; Finally, the quality control of the automatic annotation results is carried out through a two-stage verification strategy of "automatic screening+manual sampling". Dataset statistical features: Scale and division: A total of 31000 images are included, including 15500 risk images and 15500 safety images, strictly maintaining a 1:1 balanced distribution.

Divided into training set (21700 images), validation set (3100 images), and test set (6200 images) in a ratio of 7:1:2. Fine grained coverage: The system covers 5 primary risk categories and 31 secondary risk subcategories defined by national standards, and adopts a balanced sampling strategy to uniformly control each category sample to 500, effectively alleviating the problem of long tail distribution and category imbalance. This dataset fills the gap in the fine-grained AIGC risk image dataset and provides reliable data support for the objective evaluation and model training of fine-grained content security detection technology at the output end of large models.