Image Forgery / AI Image Detection System

A specialized deep learning system for detecting AI-generated images (Deepfakes) using the MesoNet-4 architecture. This project is optimized for the CIFAKE dataset (Real vs. AI).

🚀 Features

• MesoNet-4 Architecture: Compact and efficient 4-layer CNN designed for forensic analysis.
• CIFAKE Dataset Support: Built-in tools to extract and organize the CIFAKE dataset (Real/AI).
• Multiple Model Variants: Support for standard MesoNet-4 and MesoInception-4.
• Visual Workflow: Clear visualization of the detection pipeline.
• CLI Interface: Simple command-line tools for extraction, training, evaluation, and prediction.

---

🌊 System Workflow

The system follows a streamlined pipeline from raw data to prediction.

graph LR
    A[Raw Archive<br/>archive.zip] -->|Extract| B[Dataset Split<br/>Train/Test]
    B -->|Input Image<br/>256x256| C[MesoNet-4 Model]
    
    subgraph CNN Process
    C -->|Feature Extraction| D[Conv Layers]
    D -->|Classification| E[Dense Layers]
    end
    
    E -->|Sigmoid| F[Output Probability]
    F -->|Threshold > 0.5| G[AI Generated]
    F -->|Threshold < 0.5| H[Real Image]
    
    style C fill:#f9f,stroke:#333,stroke-width:2px
    style G fill:#ff9999,stroke:#333
    style H fill:#99ff99,stroke:#333

Loading

---

🧠 Model Architecture: MesoNet-4

MesoNet-4 is designed to detect "mesoscopic" properties of images—subtle artifacts introduced by AI generation processes that are invisible to the human eye but detectable by a specialized neural network.

graph TD
    Input[Input Image<br/>256x256x3] --> Conv1[Conv2D<br/>8 filters, 3x3]
    Conv1 --> Pool1[MaxPooling<br/>2x2]
    Pool1 --> Conv2[Conv2D<br/>8 filters, 5x5]
    Conv2 --> Pool2[MaxPooling<br/>2x2]
    Pool2 --> Conv3[Conv2D<br/>16 filters, 5x5]
    Conv3 --> Pool3[MaxPooling<br/>2x2]
    Pool3 --> Conv4[Conv2D<br/>16 filters, 5x5]
    Conv4 --> Pool4[MaxPooling<br/>4x4]
    
    Pool4 --> Flatten[Flatten]
    Flatten --> D1[Dense 16<br/>Dropout 0.5]
    D1 --> Output[Dense 1<br/>Sigmoid Activation]
    
    style Input fill:#e1f5fe
    style Output fill:#fff9c4

Loading

Understanding the CNN Workflow

1. Feature Extraction (Convolutional Layers)

   • The network applies a series of learnable filters (kernels) to the image.
   • Layer 1-2: Detect low-level features like edges, textures, and noise patterns (often where AI artifacts reside).
   • Layer 3-4: Combine these features to detect higher-level structures and complex artifact patterns specific to GANs or Diffusion models.
   • MaxPooling: Reduces the image size while retaining the most important features, making the model computationally efficient.

2. Classification (Dense Layers)

   • Flatten: Converts the 2D feature maps into a 1D vector.
   • Dense Layer: Analyzes the feature vector to make a decision. Dropout is used here to prevent overfitting (memorizing the training data).
   • Output Layer: Produces a single score between 0 and 1.
     • 0: High confidence it is REAL.
     • 1: High confidence it is AI/FAKE.

---

🛠️ Installation

1. Clone the repository

   git clone https://github.com/krishnab0841/Image_Forgery_Detection_Using_CNN.git
   cd Image_Forgery_Detection_Using_CNN

2. Create a virtual environment

   python -m venv venv
   # Windows
   venv\Scripts\activate
   # Linux/Mac
   source venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

---

🚦 Quick Start

1. Process Dataset

Download the CIFAKE Dataset and place your archive.zip in the project root, then run:

python main.py extract

2. Train Model

python main.py train --model_type mesonet --epochs 50 --evaluate

3. Make Predictions

# Single image
python main.py predict --model_path models/saved_models/ai_detector_final.h5 --image path/to/image.jpg

---

⚙️ Configuration

You can tweak the system in src/utils/config.py:

Parameter	Default	Description
IMG_SIZE	(256, 256)	Input resolution for the model
BATCH_SIZE	32	Number of images processed per step
EPOCHS	50	Total training iterations
LEARNING_RATE	0.001	Speed of model optimization

---

📁 Project Structure

image-forgery-detection/
├── data/
│   ├── raw/                  # CIFAKE dataset (Real/AI)
│   └── processed/            # Processed data
├── models/
│   ├── saved_models/         # Trained .h5 models
│   └── logs/                 # TensorBoard logs
├── src/
│   ├── models/               # MesoNet and CNN architectures
│   ├── training/             # Training loop and callbacks
│   ├── inference/            # Prediction logic
│   └── utils/                # Configuration and helpers
├── main.py                   # Main CLI entry point
├── extract_dataset.py        # Dataset extraction script
└── requirements.txt          # Dependencies