Dual-layer waveform domain deep learning approach for RF fingerprinting

The widespread deployment of wireless sensors and devices is enabling the rapid development of the Internet of Things. To address the privacy and the security issues of wireless transmissions, Radio Frequency (RF) fingerprinting techniques can be used to provide an additional layer of protection. With the emergence of deep learning solutions for identifying devices, we propose a pre-processing approach that generates dual-layer waveform domain images from the captured raw I/Q-samples which can be combined with either Multilayer Perceptron Neural Network (MLPNN) or Convolutional Neural Network (CNN) architectures for RF fingerprinting. The performance of the proposed approach is evaluated using over-the-air captured data, and when combined with CNN the approach was able to identify 12 Zigbee devices with an accuracy of 99% at 24 dB Signal-to-Noise Ratio (SNR), and 89% when the SNR is varied between 16 to 24 dB. The experimental results show that the proposed pre-processing approach results in reduced training time with minimal impact on the complexity of a CNN model.