~/projects

Deepfake Guard: Multi-Signal Forensic Verification

FEATURED

A dual-branch neural network combining spatial and frequency-domain analysis to detect AI-generated video manipulation, with Grad-CAM explainability and live tracking.

PythonEfficientNet-B4PyTorchOpenCVGradio
Source
Deepfake Guard: Multi-Signal Forensic Verification

Overview

A forensic deepfake detection system built as a portfolio project, designed to identify manipulation artifacts that single-branch models miss.

What it does

  • EfficientNet-B4 spatial branch + Fast Fourier Transform frequency branch in a dual-branch ensemble
  • MTCNN face tracking with adaptive frame-rate sampling for efficient video ingestion
  • Temperature scaling with LBFGS optimizer to calibrate confidence scores
  • Grad-CAM heatmaps for structural explainability
  • Deployed via OpenCV live tracking loop and Gradio dashboard

The lesson

Frequency-domain features catch artifacts that look clean spatially — the FFT branch consistently flagged faces that passed the spatial branch. Ensembling across signal types is worth the complexity.