✨ Information Preservation with Wasserstein Autoencoders: Generation Consistency and Adversarial Robustness

🗂️ Overview

This repository contains code and resources for training Wasserstein Autoencoders (WAE) to explore generation fidelity and adversarial robustness. The project focuses on concurrent density estimation using different latent space distributions and various activation functions.

⚙️ Getting Started

1️⃣ Install Dependencies

Before running any experiments, ensure you have all the required Python packages installed. You can install them with:

pip install -r requirements.txt

2️⃣ Training the Model

You can customize the training process with the following arguments:

• --groupsort: Use Groupsort activation function. Default is 0. Set to 1 to enable.
• --js: Use Jensen-Shannon Divergence. Default is 0 (MMD is selected by default). Set to 1 to enable.
• --beta: Change latent space distribution to Beta. Default is 0 (Gaussian latent space is selected).
• --exp: Change latent space distribution to Exponential. Default is 0 (Gaussian latent space is selected).
• --gauss: Opt for Gaussian Ball distribution.
• --mnist: Use MNIST dataset.

To start training, use the following command:

python3 train.py

3️⃣ Swiss Roll Experiments

To run experiments on the Swiss Roll dataset, navigate to the ./swiss_role directory and execute:

python3 code_v6.py

Ensure that dataset.py is in the same folder. Modify it as needed for your experiments.

🛡️ Robustness Testing

Robust codes for Gaussian Ball and MNIST datasets are provided in the Robust folder. These have been tested with Cauchy, Dirichlet, and Gaussian noise. To run robust experiments on the Swiss Roll dataset, navigate to ./Robust/swiss_role/ and run:

python3 robust_code.py

🧪 Experimentation

⚙️ Gaussian Ball Clusters

To experiment with the Gaussian Ball position and increase the number of clusters, modify the dataset.py file.

🔄 MMD and JS Loss Integration

We recommend setting a portion of 0.2 for MMD and JS integration with the reconstruction loss. You can adjust this ratio in config.py.

🎛️ Hyperparameter Tuning

We encourage open collaboration on hyperparameter settings. All configurations are available in config.py for easy modification and experimentation.

🧠 Model Architecture

The model.py file contains a simple dense neural network model for MNIST and Gaussian Ball reconstruction.

🧩 Robustness Ratios

In robustness testing, datasets are mixed with specific ratios:

• Gaussian Ball: See lines 127, 129 in cauchy.py.
• Dirichlet: See lines 134, 137 in dirichlet.py.

Adjust the ratios on lines 95 and 101 respectively. Similar modifications can be made for the MNIST dataset.