GAN-N-Net Project README

Overview

The GAN-N-Net repo contains code written as part of my research to provide a method for reproducing the results presented in the associated paper. The codebase includes implementation of models with and without the proposed GAN-N-Net model enhancements.

GAN-N-Net Architecture

GAN-N-Net Architecture Overview

Model Architecture Details

Generator Architecture

Layer	Details	Activation Function
Fully Connected	Output size = 32×32×16, BatchNorm	LeakyReLU (α = 1.0)
Reshape	Output shape: (32, 32, 16)	–
Conv2D	8 filters, 5×5 kernel, stride 1, same padding, BatchNorm	LeakyReLU (α = 0.3)
Conv2D	4 filters, 5×5 kernel, stride 1, same padding, BatchNorm	LeakyReLU (α = 0.3)
Conv2D	1 filter, 5×5 kernel, stride 1, same padding, BatchNorm	tanh

Discriminator Architecture

Layer	Details	Activation Function
Input	Image sample of shape (32, 32, 1)	–
Dropout	Dropout layer with rate = 0.3	–
Conv2D	32 filters, 5×5 kernel, stride 1, same padding	LeakyReLU (α = 0.3)
Dropout	Dropout layer with rate = 0.5	–
Conv2D	64 filters, 3×3 kernel, stride 1, same padding, BatchNorm	LeakyReLU (α = 0.2)
Conv2D	128 filters, 2×2 kernel, stride 1, same padding	LeakyReLU (α = 0.2)
Global Avg Pooling	Global Average Pooling 2D layer	–

Performance Visualizations

The following graphs show the validation set convergence accuracy, precision, recall, and F1-score for each dataset:

QUIC Davis	QUIC Paris-Est Créteil	Mirage22	Flash Networks

Validation set convergence accuracy, precision, recall, and F1-score for each dataset

Due to space constraints in the paper, we presented convergence results for only two datasets. Here, we provide the complete convergence visualizations for all datasets used in our research, along with the detailed architecture tables that were removed from the paper due to space limitations.

Repository Structure

• mini.py: Script to obtain results using our GAN-N-Net model.
• mini100.py: Script to obtain results without using our GAN-N-Net model.

Getting Started

To reproduce the paper results, follow these steps:

Prerequisites

Ensure you have the requairements package installed you can use requirements.txt file pip install -r requirements.txt

Installation

Clone the repository to your local machine:

git clone https://github.com/ArielCyber/GAN-N-Net.git

cd GAN-N-Net

pip install -r requirements.txt

To get results with our model:

for QUIC Paris-Est Créteil(quic_pcaps), run:

python mini.py GAN-N-Net/datasets/mini_flowpic_quic_pcaps 

for QUIC Berkeley (quic text), run:

python mini.py GAN-N-Net/datasets/mini_flowpic_quic_text

To get results without our model

For QUIC Paris-Est Créteil(quic_pcaps), run:

python mini100.py GAN-N-Net/datasets/mini_flowpic_quic_pcaps 

For QUIC Berkeley (quic text), run:

python mini100.py GAN-N-Net/datasets/mini_flowpic_quic_text

You can also set a unique name for your traning.

For example:

python mini.py GAN-N-Net/datasets/mini_flowpic_quic_pcaps try_gannet

This is a list of all available Arguments

mini.py & mini100.py accepts the following arguments to customize its behavior:

• data_dir (Required): The directory containing your training data.

• run_name (Optional): A name for this specific run of the script. This name will be used in log files. If not provided, the data folder name will be used.

• --test_split (Optional): The fraction of the data to allocate for the test set (e.g., 0.2 represents 20%). Default is 0.1 (10%).

• --val_split (Optional): The fraction of the data to allocate for the validation set. Default is 0.3 (30%).

• --batch_size (Optional): The batch size used during training. Default is 64.

• --drop_reminder (Optional): Drop the last batch if it is not full. Default is True.

• --label_rate (Optional): Rate of labels in the training data. Default is 1 (all).

• --epochs (Optional): Number of epochs to train. Default is 200.

• --train_rate (Optional): Rate of training data. Default is 1 (all).

• --save_model (Optional): Save the model after training. Default is False.

Example Usage:

python mini.py /path/to/data/ my_training_run --test_split 0.25