SceneShield: AI Fake Image Classifier

Overview

This repository contains a deep learning model for fake scene classification, utilizing data from CIDAUT AI Fake Scene Classification 2024. The notebook includes Python code for loading data, applying data augmentations, and training a model using PyTorch.

Table of Contents

1. Introduction
2. Installation
3. Data Loading
4. Data Augmentations
5. Advanced Custom Dataset Class
6. Running the Code
7. Gaussian Mixture Model (GMM)
8. Results
9. Usage
10. License

Introduction

Fake scene classification is a challenging task in computer vision. This project aims to detect fake scenes using advanced deep learning techniques and data augmentation strategies.

Installation

To set up the environment, install the required dependencies:

pip install -r requirements.txt

Data Loading

The dataset is loaded and preprocessed using PyTorch DataLoader.

from torch.utils.data import DataLoader

data_loader = DataLoader(dataset, batch_size=32, shuffle=True)

Data Augmentations

We apply various augmentations to improve model generalization:

• Random Resizing
• Horizontal Flipping
• Color Jittering
• Gaussian Noise Injection

import torchvision.transforms as transforms

transform = transforms.Compose([
    transforms.RandomResizedCrop(224),
    transforms.RandomHorizontalFlip(),
    transforms.ColorJitter(),
    transforms.ToTensor()
])

Advanced Custom Dataset Class

A custom dataset class is implemented for more control over data processing:

from torch.utils.data import Dataset

class CustomDataset(Dataset):
    def __init__(self, data, labels, transform=None):
        self.data = data
        self.labels = labels
        self.transform = transform
    
    def __getitem__(self, index):
        img, label = self.data[index], self.labels[index]
        if self.transform:
            img = self.transform(img)
        return img, label

    def __len__(self):
        return len(self.data)

Running the Code

To train the model, execute:

python train.py

Gaussian Mixture Model (GMM)

A GMM-based approach is implemented to analyze clusters of fake and real images:

from sklearn.mixture import GaussianMixture

gmm = GaussianMixture(n_components=2, random_state=42)
gmm.fit(features)

Results

The final trained model achieves 97% accuracy on the validation set.

Usage

To test a new image, run:

python predict.py --image sample.jpg

License

This project is licensed under the MIT License.

---

Note: This code is intended for educational and research purposes only and may not be optimized for production use.