Inspiration

As generative AI becomes increasingly embedded in daily life, risks of privacy loss grow. Malicious actors exploit AI to harvest personal data, create deepfakes, or mount identity-based attacks, while ordinary AI use can also expose sensitive information through prompts, uploads, or digital traces. To safeguard users, there is a pressing need to leverage AI itself as a protective layer—detecting, preventing, and mitigating privacy breaches. This includes building tools that automatically identify and redact personal identifiers, monitor for potential data leakage, and defend against misuse of online content, enabling safer digital participation in an AI-driven world.

What it does

We propose an AI-powered privacy red-teaming tool that helps users understand the hidden risks in the media they share online. When a user uploads a photo, video, or description, the system automatically extracts signals such as audio, text, and visual cues. It then simulates how an adversary might exploit this information: searching the web, linking content to social media profiles, and surfacing sensitive details (e.g., location, identity, affiliations).

Instead of directly harvesting or leaking information, our solution acts as a privacy mirror: it shows users what others could potentially infer about them from their content. By revealing these risks proactively, users are empowered to take preventive steps (e.g., blurring faces, removing location tags, editing captions) before posting.

This approach transforms generative AI from a threat into a safeguard — using the same power that enables privacy attacks to instead defend and educate users about their digital footprint.

How we built it

We used agentic AI in the google cloud space with a custom front and backend to display risks to the user using the service. Agentic AI had both Gen AI As well as Computer Vision and clever pattern recognition to filter out such pii data

Challenges we ran into

Integrating such a big system to work all together as well as the unpredictability of LLMs

Accomplishments that we're proud of

We managed to build a full self sufficient service in less than 72 hours learning through the tough experience and high stress making us better developers

What we learned

Agentic AI, Gen AI, Computer Vision

What's next for Privacy Guard

Integration to high volume apps like tiktok

Built With

  • easyocr
  • fastapi
  • ffmpeg-python
  • gemini
  • google-adk
  • google-genai
  • numpy
  • operncv
  • python-dotenv
  • soundfile
  • uvicorn
Share this project:

Updates