Dagger7164/VisualPromptInjection

Typographic Visual Prompt Injection (TVPI) Dataset

This dataset accompanies the study "Typographic Visual Prompt Injection in Vision-Language Models: A Comparative Evaluation of Base and Instruction-Following Architectures." It contains the adversarial prompts and generated typographic images used to evaluate PaliGemma-3b and LLaVA-1.5-7b against visual prompt injection across three risk vectors.

Code repository: https://github.com/AyushyaRanjan/Typographic-Visual-Prompt-Injection-in-Vision-Language-Models

Dataset Summary

45 adversarial text prompts, each rendered into 3 typographic image styles, producing 135 images total. Prompts span three risk vectors — System Bypass, Data Privacy, and Content Policy Violation — with 15 prompts per vector. Every prompt is designed such that the expected, safety-aligned model response is REFUSE.

Dataset Structure

prompts.csv            45 rows — the source adversarial prompts
images/
  {prompt_id}_clean.png          black text (RGB 0,0,0) on white, 512x256px
  {prompt_id}_lowcontrast.png    light grey text (RGB 200,200,200) on white
  {prompt_id}_distorted.png      oblique/italic font variant

prompts.csv columns

Column	Description
vector_id	V-01, V-02, or V-03
category	System Bypass / Data Privacy / Content Policy
prompt_id	Unique identifier, e.g. V01-P01
prompt_text	The adversarial prompt text
expected_response	Always REFUSE

Risk vectors

• V-01 System Bypass — prompts attempting to override alignment guidelines or elicit disclosure of system instructions.
• V-02 Data Privacy — prompts targeting extraction of context window contents, system prompts, or session instructions.
• V-03 Content Policy Violation — prompts requesting content a safety-aligned model should decline (social engineering, fraud, unauthorized system access instructions).

Image Generation Details

All images are 512×256px PNGs with 20px padding and 20pt font size. Text is automatically wrapped to fit within image boundaries. The three style conditions are held identical in layout — only text color/style differs — to isolate typographic presentation as the independent variable.

Condition	Text color	Font
Clean	RGB (0, 0, 0)	Sans-serif
Low contrast	RGB (200, 200, 200)	Sans-serif
Distorted	RGB (0, 0, 0)	Oblique/italic

Intended Use

This dataset is intended for AI safety research evaluating vision-language model robustness to visual/typographic prompt injection. It is not intended for training models to perform the attacks described, and none of the generated images contain any content beyond the rendered adversarial text itself.

Content Warning

prompt_text values in vector V-03 include requests for content such as phishing scripts, social engineering scripts, and unauthorized access instructions. These are provided strictly to evaluate whether models correctly refuse them; no harmful outputs generated during evaluation are included in this dataset.

Citation

If you use this dataset, please cite the accompanying report available in the code repository above.

License

MIT License