AI Image Detection Using Deep Learning

A deep learning-based web application for detecting whether an image is Real or AI-Generated. The project compares multiple CNN and transfer-learning architectures and integrates the best-performing model into a Django web application.

📌 Project Overview

The rapid development of generative AI has made it increasingly difficult to distinguish AI-generated images from real images. This project investigates deep learning approaches for binary image classification and develops a web-based detection system.

The project follows a complete machine learning workflow:

Dataset Acquisition → Data Cleaning → Duplicate Detection → Leakage-Aware Splitting → Preprocessing → Augmentation → Model Training → Fine-Tuning → Evaluation → Django Deployment

🎯 Objectives

• Detect AI-generated images using deep learning.
• Investigate and reduce potential data leakage.
• Compare different CNN and transfer-learning architectures.
• Evaluate models using multiple performance metrics.
• Deploy the best-performing model through a Django web application.
• Provide users with prediction results and prediction history.

🗂️ Dataset

The project uses the CIFAKE (Real and AI-Generated Synthetic Images) dataset.

Initial dataset:

• 120,000 images
• 60,000 Real
• 60,000 AI-Generated
• Image size: 32 × 32
• RGB images

Duplicate and leakage analysis was performed before model training.

After cleaning and removing duplicate-related leakage:

• 118,664 images
• Training: 78,961
• Validation: 19,741
• Test: 19,962

Exact hash overlap between the final train, validation and test splits was verified to be zero.

🔬 Methodology

1. Data Cleaning

The dataset was inspected for:

• Corrupted images
• Duplicate images
• Image dimensions and colour modes
• Class distribution
• Potential train-test leakage

MD5 hashing was used to identify exact duplicate images.

2. Preprocessing

Images were:

• Converted to RGB
• Resized to 224 × 224
• Converted to float32
• Normalized to the range 0–1

3. Data Augmentation

Training images were augmented using:

• Random horizontal flipping
• Random rotation
• Random translation
• Random zoom

This was used to reduce overfitting and improve model generalisation.

4. TensorFlow Data Pipeline

The TensorFlow pipeline performs:

Load Image
    ↓
Decode JPEG
    ↓
Convert to RGB
    ↓
Resize to 224 × 224
    ↓
Normalize to 0–1
    ↓
Apply Augmentation
    ↓
Batch (32)
    ↓
Prefetch
    ↓
Model

🧠 Models

Four architectures were investigated:

Baseline CNN

A custom CNN was developed as a baseline reference model.

Architecture includes:

• Conv2D layers
• MaxPooling
• GlobalAveragePooling
• Dense layer
• Dropout
• Sigmoid output

ResNet50

ResNet50 with ImageNet pretrained weights was used for transfer learning.

A two-stage training strategy was applied:

Stage 1

• Freeze the pretrained backbone
• Train the classification head

Stage 2

• Unfreeze deeper layers
• Fine-tune using a low learning rate

MobileNetV2

MobileNetV2 was included as a lightweight transfer-learning alternative with significantly fewer parameters than ResNet50.

EfficientNetB0

EfficientNetB0 was evaluated as another efficient transfer-learning architecture. In the tested configuration, it remained close to chance-level performance and therefore was not selected for deployment.

📊 Results

Model	Accuracy	ROC-AUC	Result
Baseline CNN	95.01%	98.80%	Baseline
ResNet50 Stage 1	96.78%	99.52%	Improved
ResNet50 Fine-Tuned	98.17%	99.76%	Selected
MobileNetV2 Stage 1	91.41%	97.39%	Lower
MobileNetV2 Fine-Tuned	96.98%	99.60%	Lightweight alternative
EfficientNetB0 Stage 1	~50.66%	50.00%	Not selected

Selected Model

Fine-tuned ResNet50 achieved the best observed validation performance:

• Accuracy: 98.17%
• Precision: 98.14%
• Recall: 98.14%
• F1 Score: 98.14%
• ROC-AUC: 99.76%

The fine-tuned ResNet50 model was therefore selected for integration into the Django application.

🌐 Web Application

The trained models are integrated into a Django web application.

Main Features

• User registration
• User login/logout
• Image upload
• AI-generated image detection
• Multiple model selection
• Confidence score
• Prediction history
• Model comparison
• About/project information

Detection Workflow

User Uploads Image
        ↓
Select Detection Model
        ↓
Django Backend
        ↓
Image Preprocessing
        ↓
Trained CNN Model
        ↓
Prediction Probability
        ↓
0.5 Threshold
        ↓
Real / AI-Generated
        ↓
Confidence Score
        ↓
Prediction History

🛠️ Technologies Used

Machine Learning

• Python
• TensorFlow
• Keras
• NumPy
• PIL
• Scikit-learn

Deep Learning

• CNN
• Transfer Learning
• ResNet50
• MobileNetV2
• EfficientNetB0
• ImageNet pretrained models

Web Development

• Django
• HTML
• CSS
• SQLite
• Django ORM

Development Environment

• Jupyter Notebook
• Google Colab
• Visual Studio Code
• Git
• GitHub

📁 Project Structure

AI-Image-Detection-Deep-Learning/
│
├── aiImageDetector/
│   └── Django project configuration
│
├── detector/
│   ├── ai_models.py
│   ├── forms.py
│   ├── models.py
│   ├── urls.py
│   └── views.py
│
├── ml_models/
│   ├── baseline_cnn_best.keras
│   ├── mobilenetv2_finetuned_best.keras
│   ├── mobilenetv2_stage1_best.keras
│   ├── resnet50_finetuned_best.h5
│   ├── resnet50_finetuned_best.keras
│   └── resnet50_stage1_best.keras
│
├── notebook/
│   └── 3190951_AI Image Detection.ipynb
│
├── static/
├── templates/
├── manage.py
├── requirements.txt
└── .gitignore

🚀 Running the Web Application

1. Clone the repository

git clone https://github.com/19jenil/AI-Image-Detection-Deep-Learning.git
cd AI-Image-Detection-Deep-Learning

2. Create a virtual environment

Windows:

python -m venv venv
venv\Scripts\activate

macOS/Linux:

python3 -m venv venv
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Apply database migrations

python manage.py migrate

5. Start the Django server

python manage.py runserver

Then open:

http://127.0.0.1:8000/

📓 Research Notebook

The complete machine learning experimentation workflow is available in:

notebook/3190951_AI Image Detection.ipynb

The notebook contains dataset preparation, leakage analysis, preprocessing, augmentation, model training, fine-tuning and evaluation.

⚠️ Disclaimer

The reported performance represents the results obtained on the selected CIFAKE dataset and experimental configuration. It should not be interpreted as guaranteed performance on all real-world images or unseen generative AI systems.

👨‍💻 Author

Jenil Patel