covert-ml is a GAN-based covert communication method that enables establishing a reliable, undetectable covert channel within autoencoder wireless communication systems with the minimum impact.
This repository contains the project that was created throughout the course of my master's thesis, titled "Covert Communication in Autoencoder Wireless Systems." This research project explores the concept of using generative AI to produce covert signals within a Generative Adversarial Network (GAN) framework. The objective is to create signals that are indistinguishable from normal signals, enabling covert communication in autoencoder wireless systems.
To run this project locally, follow these installation steps:
- Clone this repository to your local machine.
git clone https://github.com/ali-mohammadi/covert-ml.git- Install the required Python packages and dependencies
Once the installation is complete, you can use the project for training the autoencoder, covert models, and evaluating their results. The main scripts for each step are provided in the project repository. Here's an overview of how to use the project:
-
Autoencoder Training/Evaluation: Execute the
wireless_autoencoder/siso/train.pyorwireless_autoencoder/mimo/train.pyscripts to train the autoencoder using the prepared dataset. Model hyperparameters can be modified fromparameters.pyas needed for your specific use case. -
Covert Models Training/Evaluation: After successfully training the autoencoder, use the
wireless_covert/siso/train.pyorwireless_covert/mimo/train.pyscripts to train the covert models. Similarly, hyperparameters can be modified fromparameters.pyas needed for your specific use case.
You can also change the channel configuration from shared_parameters.py
Covert and autoencoder wireless models were trained on a high- performance desktop computer with an RTX 3080Ti graphics card. The desktop configuration included an Intel Core i7 10700KF 3.80GHz CPU, 16GB of DDR4 3200MHz RAM, and 12GB of GPU memory.
Although training the models using a high performance GPU significantly reduce the training time, we have provided support for both CPU and GPU training in this project.
The broadcast nature of wireless communications presents security and privacy challenges. Covert communication is a wireless security practice that focuses on intentionally hiding transmitted information. Recently, wireless systems have experienced significant growth, including the emergence of autoencoder-based models. These models, like other DNN architectures, are vulnerable to adversarial attacks, highlighting the need to study their susceptibility to covert communication. While there is ample research on covert communication in traditional wireless systems, the investigation of autoencoder wireless systems remains scarce. Furthermore, many existing covert methods are either detectable analytically or difficult to adapt to diverse wireless systems. The first part of this thesis provides a comprehensive examination of autoencoder-based communication systems in various scenarios and channel conditions. It begins with an introduction to autoencoder communication systems, followed by a detailed discussion of our own implementation and evaluation results. This serves as a solid foundation for the subsequent part of the thesis, where we propose a GAN-based covert communication model. By treating the covert sender, covert receiver, and observer as generator, decoder, and discriminator neural networks, respectively, we conduct joint training in an adversarial setting to develop a covert communication scheme that can be integrated into any normal autoencoder. Our proposal minimizes the impact on ongoing normal communication, addressing previous works shortcomings. We also introduce a training algorithm that allows for the desired tradeoff between covertness and reliability. Numerical results demonstrate the establishment of a reliable and undetectable channel between covert users, regardless of the cover signal or channel condition, with minimal disruption to the normal system operation.
Our covert model successfully demonstrated its ability to embed secret messages into covert signals without relying on handcrafted features. Through the utilization of the generative adversarial training framework, we significantly reduced the detection probability of the covert signals produced. Furthermore, we proposed a training procedure that allows us to adjust the trade-off between the conflicting objectives of covert communication, which are reliability of communication and the probability of detection. This adjustment is achieved by introducing regularizers into the model’s loss function, independent of channel conditions or user messages. Our findings demonstrate that our covert model is channel-agnostic and insensitive to cover signals. To evaluate the performance of our model, we conducted assessments across three channel models: AWGN, Rayleigh, and Rician fading. We varied covert rates and the number of system users to analyze the model’s robustness. Additionally, we investigated the impact of our covert signals on the ongoing normal system and confirmed that our covert scheme causes minimal disruption to the system.
Covert Communication in Autoencoder Wireless Systems (Ali Mohammadi Teshnizi, Majid Ghaderi, Dennis Goeckel)
If you have any questions, suggestions, or inquiries related to this project or thesis, feel free to contact me at:
- Ali Mohammadi: ali.mohammaditeshniz@ucalgary.ca