Inspiration

Generative AI has unlocked new possibilities but also new risks: with just a single photo, powerful models can infer who you are and where you are. This creates urgent concerns around identity theft, stalking, and location-based targeting. Our solution responds directly to the hackathon’s call to use AI to defend user privacy.

What it does

We developed a two-part privacy toolkit:

GeoShield detects and obscures location-revealing cues in images, such as license plates, road signs, or logos. Using AI models (Google Vision, YOLO, CLIP) combined with LaMa inpainting, it masks sensitive regions while keeping the photo natural and shareable.

FaceShield is a privacy-first photo gallery that automatically blurs faces before display or sharing. Users can safely manage their photos, adjust blur levels, unlock specific images with a password, and rest assured that all EXIF/GPS metadata is stripped.

How we built it

Our project was built with Python, React, Tailwind CSS, OpenCV, FastAPI, and Google Cloud Vision API, alongside lightweight ML libraries for detection and inpainting. Assets include curated photo datasets for testing privacy scenarios.

Challenges we ran into

We initially found it challenging to define a clear use case for a GenAI-driven privacy protection tool for consumers or TikTok users like ourselves. A natural question that arose was: why would creators want to hide their face or location if they chose to post content publicly in the first place?

On further reflection, we concluded that our tool could serve as a risk detector for situations where creators unintentionally expose personal details, such as visual clues that may reveal their residence or location, potentially compromising their safety.

Another application would be enabling creators to control access to original images. For instance, if a post contains faces of individuals who do not wish to be publicly visible, creators could restrict the unedited version to mutual friends. Public viewers would instead see either a blurred or censored version, or a generative mask (e.g. a cartoon overlay) that preserves the overall user experience of scrolling through posts. This represents a promising extension we could explore further if given more time.

Another challenge we encountered was determining which features are critical for identifying a landmark or location. How do we strike the balance between reconstructing the landmark for risk detection and preserving the visual integrity of the original image the creator intended to share? For instance, simply introducing noise into the background is neither visually appealing to viewers nor aligned with what creators want. A more practical balance is to selectively remove identifiable elements such as text or logos, or make subtle alterations like changing building tiles or colors. In the future, we could further refine this approach by incorporating user feedback and offering creators greater autonomy over how their content is protected.

Extensions

Beyond the extension mentioned earlier, our technology could also be applied to videos, though this would require more time given the higher processing load from frame rates. Additionally, we could integrate NLP analysis of captions to enable a comprehensive multi-modal risk assessment of a creator’s post.

Conclusion and Submission Materials

By focusing on AI for Privacy, GeoShield & FaceShield demonstrate how intelligent detection, local-first processing, and intuitive controls can give users back ownership of their digital privacy.

GitHub Repository: github.com/jianronggu/The-Power-Puff-Girls

Demo Video (YouTube): https://youtu.be/yMuwkrj_k44?si=ZVUeaK2QbtOpt8W5