The problem FaceFortress: Privacy-Shielded Deepfake Detection solves

This project tackles the growing problem of deepfakes by creating a user-friendly tool to spot fake images. It uses a smart AI model to analyze faces and decide if they're real or not. What's cool is that it doesn't just give you a yes or no answer - it shows you which parts of the image it focused on to make its decision.
The really innovative part is how it handles privacy. Using some fancy tech called Nillion, it can do its calculations without actually seeing or exposing the sensitive parts of the images. This means different organizations can work together to improve deepfake detection without worrying about data privacy issues.
You just upload a photo through a simple web interface, and it tells you if it thinks the image is real or fake, along with how confident it is. It's a practical solution that could be super useful for fighting misinformation or verifying images in important situations.

Challenges we ran into

Sure, I'll highlight some of the challenges you faced, including the computational limitations and your interactions with the Nillion team. Here's a human-like description of the challenges:
We hit a few roadblocks while building this deepfake detector. The biggest headache was the computation time - it was taking way too long to process images, which isn't great for a tool that needs to be quick and responsive.
Another major issue was that the system could only handle a small number of parameters. This limited how detailed our analysis could be, potentially affecting the accuracy of our deepfake detection.
We reached out to the Nillion team about these problems. They suggested we needed to reduce the complexity of our computations to make the system more efficient. This meant we had to rethink our approach and find ways to simplify our model without sacrificing too much accuracy.
Balancing privacy with performance was tricky too. While Nillion's tech is great for keeping data secure, it added another layer of complexity to our already resource-intensive process.
We also struggled with integrating all the different parts - the face detection, the deepfake analysis, and the secure computation. Getting everything to work together smoothly was like solving a puzzle with pieces that didn't quite fit at first.
Despite these challenges, we kept tweaking and optimizing, and eventually got to a version that works pretty well, even if it's not perfect yet.