DeepGuard: Deepfake Face Image Detector

A Jupyter notebook that trains an EfficientNet-B0 based binary classifier to distinguish real human face photographs from AI-generated/deepfake faces, with explainable forensic analysis for each prediction.

Features

• Transfer learning from ImageNet-pretrained EfficientNet-B0
• Two-phase training: frozen backbone warm-up, then full fine-tuning with early stopping
• Explainable analysis: Grad-CAM heatmaps, frequency spectrum (FFT), and forensic evidence
• Fused detection: combines CNN predictions with handcrafted forensic cues (noise uniformity, texture complexity, edge gradients, symmetry)
• Supports CUDA, MPS (Apple Silicon), and CPU

Example Output

Full Forensic Analysis

Grad-CAM (Model Attention)

Frequency Spectrum (FFT)

Detection Scores

Input Image & Forensic Report

In this example, the model classifies an AI-generated image as FAKE (66.5% fused verdict). The forensic analysis explains why:

• Noise is very uniform across the image — characteristic of AI-generated content
• Edge gradients are unusually uniform — real faces show more variation across eyes, nose, jawline

Training Outputs

Getting Started

Prerequisites

• Python 3.10+
• PyTorch, torchvision, timm, scikit-learn, matplotlib, seaborn, Pillow, ipywidgets, opencv-python

Setup

pip install torch torchvision timm scikit-learn seaborn matplotlib pillow ipywidgets opencv-python