🧠 TrueSight_AI – Multi-Modal Deepfake Detection System

RGB + Noise + Frequency Feature Fusion for Real vs Fake Image Classification

##If you want dataset for this please download

🔍 TrueSight_AI — Multi-Modal Deepfake Detection

TrueSight_AI is a three-branch deepfake detection system that analyzes:

• RGB domain
• Noise (PRNU) domain
• Frequency (FFT) domain

This multi-modal design catches deepfakes far more reliably than RGB alone.

---

📊 System Flowchart

`flowchart.png`

📘 1. Overview

Deepfake detection is hard using only RGB pixels.

Different domains reveal different hidden clues:

Domain	What It Detects	Why It Works
RGB	Visual noise, textures, shadows	GANs hallucinate micro-details
Noise	Missing PRNU camera fingerprints	AI images lack real sensor noise
Frequency	Checkerboard patterns, FFT spikes	GANs leave periodic artifacts

TrueSight_AI extracts all three and merges them into a unified classifier.

---

🧬 2. Theory Behind Each Branch

🔴 RGB Branch

A CNN extracts:

• Texture details
• Color patterns
• Edge behavior
• Micro-structure inconsistencies

GANs often miss tiny high-frequency structures.

---

🟣 Noise Branch (PRNU Residuals)

Noise is extracted using: Noise = Image - GaussianBlur(Image)

This branch detects:

• Missing camera sensor noise
• Over-smoothness
• Denoising artifacts
• Fake PRNU patterns

Real images always contain PRNU → AI images don’t.

---

🔵 Frequency Branch (FFT Features)

FFT reveals:

• High-frequency spikes
• Checkerboard artifacts
• Moiré patterns
• Periodic GAN noise

These are invisible in RGB, but obvious in frequency space.

---

🏗 3. Project Structure

🏗 3. Project Structure go Copy code TrueSight_AI/ │── model.py │── train.py │── test_model.py │── predict_single.py │── verify_setup.py │── dataset.py │── feature_extractor.py │── saved/ │ ├── best_model.pth │ ├── training_curves.png │ ├── confusion_matrix.png │── dataset/ ├── train/ │ ├── real/ │ └── fake/ └── test/ ├── real/ └── fake/

---

⚙️ 4. Installation

pip install torch torchvision opencv-python pillow numpy matplotlib tqdm seaborn scikit-learn

🧪 5. Verify Setup

Run:

python verify_setup.py

This script verifies that your environment is ready by checking:

📁 Dataset paths 🔧 Model imports ⚡ GPU / MPS / CPU availability 🖼️ Image loader integrity

🏋️ 6. Train the Model

python train.py best_model.pth training_curves.png confusion_matrix.png metrics.json

🧾 7. Evaluate Full Test Dataset

bash Copy code python test_model.py Outputs:

Accuracy Precision Recall F1-Score Confusion matrix Wrong predictions

📊 8. Generated Outputs

Training curve Confusion matrix Real vs Fake confidence Per-image prediction logs

🚀 9. Future Enhancements

Add Vision Transformer (ViT) branch Add Wavelet Transform (DWT) Integrate CLIP features Deploy using FastAPI or Streamlit