Academic Project Grant Request

Type of project

Academic project

Project

DeepFake Detector
Deployment: https://deepfake-detector.app

This is an academic project developed under the MacAI Society at McMaster University. Our goal is to build an accessible, research-grounded tool for detecting AI-generated and manipulated images, while improving public understanding of model confidence and explainability.

Why this project matters

Deepfakes are becoming easier to generate and harder to identify, which creates risks for misinformation, identity abuse, and trust in visual media. We are building this project as an open-science effort to:

• Provide a practical, student-led detection system for real users
• Compare model behaviors and failure modes in a transparent way
• Share learnings, methods, and implementation details with the community

Brief architecture

The system uses a production web architecture with separate frontend and backend services:

• Frontend: TypeScript/React web app (public interface at deepfake-detector.app)
• Backend API: FastAPI service deployed on Hugging Face Spaces GPU
• Inference stack: Multi-model ensemble (CNN + ViT/DeiT + gradient-field CNN)
• Fusion layer: Logistic-regression stacking to combine submodel probabilities
• Explainability: Heatmaps/attention outputs plus optional LLM-based interpretation
• Model hosting/cache: Hugging Face Hub with persistent cache on mounted storage

Grant use

Additional HF resources would be used to improve reliability and research velocity:

• Faster cold starts and more stable GPU-backed deployments
• Reproducible evaluation across larger image sets
• Better support for open demos, student workshops, and public testing

We are happy to keep this work open and share progress/results with the community.