← Aman Kumar Limited · RegisterNotes to the accounts · DFAK
Featured · research · 2024 · Shipped
Indian Deepfake Image Detection
A deepfake detector tuned for Indian faces — 99.96% accuracy on 140K images, served through a React + FastAPI web app.
- Role
- Solo — model, pipeline, web app
- Solo
- Outcome
- A deepfake detector tuned for Indian faces — 99.96% accuracy on 140K images, served through a React + FastAPI web app.
- Links
- Code ↗
- Stack
- PyTorch · XceptionNet · MTCNN · YOLOv11 · FastAPI · React
Problem
Off-the-shelf deepfake detectors are trained mostly on Western faces. They miss more on Indian faces.
What I built
- A curated dataset of 140,000 images: 70K real and 70K synthetic.
- Automated face extraction with MTCNN and YOLOv11, plus heavy augmentation.
- An XceptionNet-based classifier, compared against custom CNNs.
- A React front end and FastAPI back end for real-time inference.
Result
99.96%Accuracysrc: resume 2025The model reached 99.96% accuracy with a 100% true-negative rate on fake detection.
This is a portfolio, not an offer of securities. No shares exist. Figures are about shipped work.