Adversarial Prompt and Annotated Image Dataset for Bias Analysis in Text-to-Image Models

Paper 1: Towards an Adversarial-Attack Approach to Assess Bias in Text-to-Image Models

This dataset accompanies the paper and includes the following resources:

biased-prompts-list.xlsx – List of all biased prompts used for image generation.

generated-images.zip – Archive containing all generated image files.

annotated-images-dataset.xlsx – Spreadsheet with annotations for each generated image.

GPT-4 Log Execution-Annotation.docx – Log of GPT-4 executions showing the annotation process and corresponding outputs.

spreadsheets-for-results-analysis.zip – Folder containing the spreadsheets used for quantitative results analysis.

The associated Jupyter Notebooks (available at https://github.com/ufal-waterloo/biased-prompts) provide scripts to:

Annotate images automatically using GPT.

Generate images via Hugging Face API calls using the prompt lists.

Explore the annotated dataset to compute bias-related metrics across gender and skin tone.

All required files from this dataset (prompt lists, generation scripts, and annotated datasets) should be placed inside the resources/ folder in the github repository to execute the notebooks properly.

论文1：面向对抗攻击方法的文本生成图像模型偏见评估研究

本数据集随本论文同步发布，包含以下资源：
biased-prompts-list.xlsx：用于图像生成的全部偏见提示词（prompt）列表
generated-images.zip：包含所有生成图像文件的压缩存档
annotated-images-dataset.xlsx：收录每张生成图像标注信息的电子表格
GPT-4 Log Execution-Annotation.docx：记录GPT-4执行流程的日志文档，展示了标注过程及对应输出结果
spreadsheets-for-results-analysis.zip：包含用于定量结果分析的电子表格的压缩存档

关联的Jupyter笔记本（Jupyter Notebooks，可从https://github.com/ufal-waterloo/biased-prompts获取）提供以下脚本功能：
- 基于GPT自动完成图像标注
- 依托提示词列表，通过Hugging Face API调用生成图像
- 浏览标注数据集，计算与性别、肤色相关的偏见评估指标

为确保Jupyter笔记本能够正常运行，本数据集的全部必要文件（提示词列表、生成脚本及标注数据集）需放置于GitHub仓库内的resources/文件夹中。