🖼️ AI Generated Image Detector

Detection of AI-generated vs real images using deep learning with multi-domain feature representations.

This project explores different deep learning approaches for detecting AI-generated images and compares their performance.
The final system combines RGB features, gradient artifacts, and frequency-domain information with a ResNet18 backbone, achieving 95.05% accuracy on a subset of the CIFAKE dataset.

---

📌 Overview

Recent generative models such as GANs and diffusion-based models can produce highly realistic synthetic images. Detecting such images has become an important task in:

• 🔍 Digital forensics
• 📰 Misinformation detection
• 🤖 AI safety and authenticity verification

This project investigates three different approaches:

1. CNN trained from scratch
2. Transfer learning using ResNet18
3. Multi-domain feature architecture (RGB + Gradient + FFT) with ResNet18

The final approach significantly improves performance by combining spatial, edge, and frequency-domain information.

---

📂 Dataset

This project uses a subset of the CIFAKE dataset.

🔗 Dataset source:
https://www.kaggle.com/datasets/birdy654/cifake-real-and-ai-generated-synthetic-images

Dataset Subset Generation

The original CIFAKE dataset contains a much larger number of images.
Since training on the full dataset can be time-consuming on CPU, a smaller subset was created for experimentation and faster iteration.

A custom script create_subset.py was used to randomly sample images from the original dataset and generate a smaller training and testing split.

Dataset Structure

data
│
├── train_small
│ ├── real
│ └── fake
│
└── test_small
├── real
└── fake

Dataset Size

Training subset

• Real images: 5000
• Fake images: 5000

Test subset

• Real images: 2000
• Fake images: 2000

---

🧠 Methods Compared

Three different architectures were trained and evaluated.

---

1️⃣ CNN (Trained from Scratch)

A simple convolutional neural network was trained directly on gradient images.

Result

Accuracy: 76%

Limitations

• Limited feature representation
• Training from scratch requires more data
• Struggles to capture complex patterns in generated images

---

2️⃣ Transfer Learning (ResNet18)

A pretrained ResNet18 model was used with transfer learning.

Modifications

• Final classification layer modified for binary classification

Result

Accuracy: 87%

Benefits

• Pretrained features capture general visual patterns
• Faster convergence compared to training from scratch

---

3️⃣ Multi-Domain Architecture (RGB + Gradient + FFT)

The final approach combines three complementary feature domains:

Feature	Purpose
RGB	Captures textures and color information
Gradient	Detects structural artifacts using Sobel edges
FFT (Frequency domain)	Reveals periodic artifacts introduced by generative models

Input Representation

[R, G, B, Gradient, FFT]

Input shape:

5 × 224 × 224

The combined tensor is fed into ResNet18 with a modified first convolution layer to accept 5 input channels.

Result

Accuracy: 95.05%

This demonstrates that combining multiple domains significantly improves detection performance.

---

⚙️ Training Configuration

Parameter	Value
Backbone	ResNet18
Input Size	224 × 224
Batch Size	8
Optimizer	Adam
Learning Rate	0.0001
Epochs	10
Device	CPU

---

📊 Performance Comparison

Model	Accuracy
CNN (from scratch)	76%
ResNet18 Transfer Learning	87%
RGB + Gradient + FFT + ResNet18	95.05%

---

🔍 Explainability

The project also includes Grad-CAM visualization to understand which regions influence the model’s predictions.

Grad-CAM highlights regions that contain patterns typical of AI-generated artifacts.

Grad-CAM visualization showing model attention for real and AI-generated images.
The model focuses on artifact-prone regions in generated images while showing more distributed attention in real images.

---

📈 Evaluation Metrics

The evaluation pipeline produces several metrics to assess the model's ability to distinguish between real and AI-generated images.

---

Confusion Matrix

The confusion matrix shows the number of correct and incorrect predictions made by the model.

---

Precision–Recall Curve

The Precision–Recall curve highlights the trade-off between precision and recall across different thresholds.

---

ROC Curve

The ROC (Receiver Operating Characteristic) curve shows the model's performance across different classification thresholds.

---

📁 Project Structure

AI_Image_Classifier
│
├── data
|   |── train
|   |── test
|   |
│   ├── train_small
|   |   ├── real
|   |   └── fake
|   |
│   └── test_small
|       ├── real
|       └── fake
│
├── src
|   ├── create_subset.py
│   ├── dataset.py
│   ├── model.py
│   ├── train.py
│   ├── evaluate.py
│   └── gradcam.py
│
├── best_model_rgb_grad_fft.pth
└── requirements.txt

---

🚀 Installation

Clone the repository:

git clone https://github.com/yourusername/ai-generated-image-detector.git
cd ai-generated-image-detector
Clone the repository:

Install dependencies:

pip install -r requirements.txt

🏋️Training

Run training:

cd src
python train.py

📊Evaluation

To run evaluation metrics:

python evaluate.py

🔬Grad-CAM Visualization

To visualize model attention:

python gradcam.py

This produces heatmaps showing which regions influenced the prediction.

🛠️Technologies Used

• 🐍 Python
• 🔥 PyTorch
• 📷 OpenCV
• 🔢 NumPy
• 📊 Matplotlib
• 🤖 Scikit-learn
• 🎯 TorchCAM

---

👨‍💻 Author

This project was developed to explore deep learning techniques for detecting AI-generated images using multi-domain feature representations.