Use of Machine Learning for Radio Modulation Classification
August 2022 (277 Words, 2 Minutes)
Under the supervision of Professor CHEN He, this project explores the use of machine learning and deep learning techniques, specifically Convolutional Neural Networks (CNN), for radio modulation classification. The primary goal is to verify the identity of transmitters and assist in authorization systems. The approach involves comparing the efficiency of radio modulation classification using learned features against expert features-based methods on Google Colab. The advantage of using deep learning techniques is their ability to operate on raw I/Q samples and identify devices based on unique hardware-induced signal modifications.
In the context of smart cities, autonomous vehicles, the Internet of Things (IoT), and wirelessly connected devices, the ability to reconfigure systems and protocols within the communication architecture is essential. With this increasing complexity, securing these devices becomes more challenging. Traditional cryptography or security-based methods for authentication are not suitable for many IoT devices, as their performance depends on the receiver quality and requires manual feature engineering.
Machine learning techniques, especially deep learning, have demonstrated great success in image and speech recognition. Their potential for device-level fingerprinting through feature learning has been gradually shown. Deep learning approaches, such as CNN, are more robust and can extract better features from signals, leading to higher accuracy compared to feature-based approaches. By serving complex data as a dimension of two real-valued in-phase and quadrature (I/Q) inputs to the CNN, the approach is based on strategies from image and voice recognition domains in machine learning.
The main idea behind radio frequency (RF) fingerprinting is to extract unique patterns or features and use them as signatures to identify devices. In this research project, the focus is on studying those features inherent to a device’s hardware, unchanging, and not easily replicated by malicious agents. Deep learning offers a useful framework for supervised learning, enabling the extraction of features across a wide range of tasks and demonstrating increased accuracy in classification among current approaches.
The proposed method consists of two stages: a training stage and an identification stage. In the training stage, the CNN is trained using raw I/Q samples to solve a multi-class classification problem. In the identification stage, raw I/Q samples of the unknown transmitter are fed to the trained neural network, and the transmitter is identified based on the observed value at the output layer.
This project’s potential impact includes developing a more robust and accurate method for radio modulation classification and transmitter authentication, benefiting a wide range of wireless communication systems and IoT devices.
can also click here for report