GANs_VAEs_Phishing_Detection

Implementations of three distinct machine learning tasks GANs, VAEs, and Phishing Detection

1. Generative Adversarial Networks (GANs) for image generation.
2. Variational Autoencoders (VAEs) for both image generation and latent space analysis.
3. Phishing Detection using VAEs, which applies anomaly detection techniques to identify phishing URLs.

Additionally, the repository includes a detailed documentation manual with explanations and related questions.

Project Structure

.
├── data_phishing.csv          # Phishing dataset used for training the VAE model
├── AI_GenAI_Manual.pdf        # Manual containing questions and explanations
├── 21L_6225_A1_GenAI.ipynb    # Jupyter Notebook containing the implementation
├── 21L_6225_A1_GenAI.pdf      # Detailed assignment documentation

Tasks Covered

1. Generative Models

(a) Generative Adversarial Networks (GANs)

• Consists of a Generator and a Discriminator.
• The Generator learns to create realistic images from random noise.
• The Discriminator distinguishes between real and generated images.
• Used for generating images from MNIST, FashionMNIST, and Digits datasets.
• Evaluated using loss curves and image quality comparisons.

(b) Variational Autoencoders (VAEs)

• Uses an encoder-decoder architecture.
• The Encoder maps input data to a probabilistic latent space.
• The Decoder reconstructs inputs from the latent representation.
• Uses PCA for latent space visualization.
• Applied to MNIST, FashionMNIST, and Digits datasets.

2. Phishing Detection using Variational Autoencoders (VAEs)

• Uses an anomaly detection approach.
• The VAE is trained on legitimate URLs, learning to reconstruct them.
• Phishing URLs are detected based on high reconstruction errors.
• Preprocessing includes feature extraction and standardization.
• Achieved 70% accuracy based on reconstruction error thresholds.

Methodology

1. Data Preprocessing

• Image Datasets (MNIST, FashionMNIST, Digits)
  • Normalized pixel values for stable training.
  • Used PCA to visualize latent space representation.
• Phishing Dataset (data_phishing.csv)
  • Extracted relevant features from URLs.
  • Applied StandardScaler for normalization.
  • Divided dataset into training and validation sets.

2. GAN Architecture

• Generator: Converts random noise into realistic images.
• Discriminator: Determines real vs. fake images.
• Loss Function: Binary cross-entropy with adversarial training.
• Evaluation: Image quality and loss curve analysis.

3. VAE Architecture

• Encoder: Compresses input data into a latent space representation.
• Decoder: Reconstructs the original input from latent variables.
• Loss Function: Reconstruction loss + KL divergence.
• Evaluation: Reconstruction quality, latent space visualization.

4. Phishing Detection Model

• VAE-based anomaly detection.
• Trained only on legitimate URLs to reconstruct normal patterns.
• Anomalies (phishing URLs) identified based on high reconstruction errors.
• Evaluation metrics: ROC curve, classification report.

Results

1. Image Generation (GANs & VAEs)

• GANs produced sharper images.
• VAEs generated images with more structured latent space representations.
• Latent space analysis showed clear clustering patterns in VAEs.
• Datasets used: MNIST, FashionMNIST, Digits.

2. Phishing Detection

• ROC curve analysis demonstrated performance trade-offs.
• Achieved 70% accuracy, balancing false positives and false negatives.
• Challenges: Handling imbalanced datasets, improving feature extraction.

Running the Code

Requirements

• Python 3.8+
• Jupyter Notebook / Google Colab / VS Code
• PyTorch
• NumPy
• Pandas
• Scikit-learn
• Matplotlib

How to Use

1. Open 21L_6225_A1_GenAI.ipynb in Jupyter Notebook, Google Colab, or VS Code.
2. Run the notebook cells in sequence.
3. Visualize results for GANs, VAEs, and phishing detection.

Future Work

• Improve phishing detection by integrating additional URL-based features (e.g., SSL certificates, domain age).
• Explore advanced generative models such as Wasserstein GANs (WGANs) and β-VAEs.
• Enhance dataset diversity for more robust anomaly detection models.

Connect with Me