As AI systems are increasingly used for content moderation, productivity tools, and personal assistants, they often process user screenshots and photos. These may contain privacy-sensitive information such as OTP setup QR codes, credit card numbers, passwords written on sticky notes, or personal addresses. Detecting such content (preferably on device) before it is shared on social media or uploaded to websites is an important step for privacy and security.
This project will investigate whether current open source and lightweight AI models can reliably detect privacy-sensitive content in images. The work involves building a small benchmark dataset of staged or synthetic examples, defining categories of sensitive information, and evaluating state-of-the-art vision-language models against this dataset.
Project Goal:
The goal is to create a benchmark framework for detecting privacy-sensitive content in images and to compare the effectiveness of existing AI models in identifying different categories of private information.
Your work will involve:
Conducting a literature review on computer vision for content moderation and privacy detection.
Designing a taxonomy of sensitive categories (e.g., QR codes, passwords, IDs, financial information).
Building a dataset of synthetic or staged images representing these categories.
Running benchmark experiments with open-source vision-language models (e.g., LLaVA, Qwen-VL) and comparing their performance.
Summarizing strengths, weaknesses, and gaps in current AI detection methods.