Inspiration

Concealed was inspired by the prevalence of deepfakes, unconsented scraping for training ai models, and a need for a tool for better privacy in a time where everything is easily accessible online for anyone or anything to see.

What it does

This project uses a neural network that adds noise to an image to confuse ai models trying to see what's depicted. This helps poison future datasets and prevents ai from modifying images, but is subtle enough that very little quality is lost, not enough to really effect human perception.

How we built it

We used fastapi with a react frontend to run the application. We also used google antigravity to assist our development, and 4 snowflake small gpu instances to train our neural network.

Challenges we ran into

The neural network we were training degraded the image quality beyond a subtle overlay, and we needed to reset to a previous version, keeping simple instructions, to make sure the network stayed on track.

Accomplishments that we're proud of

We trained a neural network using snowflake, we worked together well as a team, and we're pretty excited about our concept.

What we learned

Keep it super simple! Overcomplicating a project can lead to difficulty implementing features and large mistakes.

What's next for Concealed

Since we used a researched algorithm to generate noise, with more training our neural network will be able to conceal images from any model, open or closed source.

Built With

Share this project:

Updates

Submission history