Fusion-AI Deepfake Detection

Inspiration

AI-generated face images are increasingly misused for misinformation and fraud. This student project explores practical deepfake-image screening using the MesoNet family of compact detectors.

What it does

Takes a face image and returns a real/fake prediction with a confidence score. The repository keeps training/inference code (see MesoNetDetect); the earlier Vercel demo is currently offline so the code runs locally.

How I built it

• Python + MesoNet-style CNN detector (see MesoNetDetect in the repo)
• Training experiments in notebooks/scripts in the repo
• Web UI deployable on Vercel

Challenges

Generalizing across generators and resolutions with a small model, and keeping inference fast enough for a web demo.

What I learned

Compact CNN design trade-offs, evaluation on mixed real/fake sets, and shipping an ML demo end to end.

Built for

ML Empowerment Build Challenge 3.0 - AI social-good track (trust & safety tooling).

Links

• Code: https://github.com/Priyanshu85277/Fusion-AI

Built With

• computer-vision
• deep-learning
• machine-learning
• python
• vercel