DeepGuard: AI Deepfake Detector

Checking video and audio for signs of manipulation, with three independent detectors.

DeepGuard analyses video and audio files to spot synthetic media. It combines CNN-based visual analysis, spectral audio processing and 468-point facial landmark tracking into a single authenticity score, and is paired with an educational chatbot that explains what it finds.

The problem

Synthetic media is getting harder to spot by eye, and a single check is easy to fool. DeepGuard looks at the file from three angles at once and shows how it reached its verdict instead of just returning a yes or no.

What I built

• Video analysis: CNN-based detection of visual artifacts, compression anomalies and colour inconsistencies.
• Audio analysis: Spectral analysis using MFCC features to detect synthetic speech.
• Face landmarks: A 468-point facial mesh used to look at blink patterns, lip-sync and micro-expressions.
• Confidence scoring: One combined authenticity score, with a breakdown of each detector's contribution.
• Privacy first: Uploaded files are deleted immediately after analysis.

How it is put together

A FastAPI backend keeps each detector in its own module (video, audio and face analysis) so they can be improved separately. The interface is plain HTML, CSS and JavaScript hosted on Vercel, and the API runs on Render.

Current limits

The backend runs on a free tier, so the first request after a quiet period can take 50 to 60 seconds while it wakes up.

Built with

Python, FastAPI, TensorFlow, JavaScript, Vercel, Render

Source on GitHub · Live demo

All projects