Attention is Needed for RF Fingerprinting

Radio Frequency (RF) fingerprinting is a novel solution for identifying a unique radio from a large pool of devices by analyzing the subtle characteristics that are inherent in the radio waveform. Deep convolutional neural networks have been widely used to handle the RF fingerprinting task because of their exceptional capacity for representation learning. However, there are still challenges in employing deep convolutional neural networks, such as how to enable the model learn more robust and discriminative RF fingerprints. This paper aims to explore new model architectures to learn robust RF fingerprints. Hence we proposes a novel Dual Attention Convolution module that simultaneously learns channel attention and spatial attention to tune the RF fingerprints, enhancing the convolutional layers’ potential for representation learning. Our proposed module is lightweight and plug-and-play. A number of convolutional neural networks can be equipped with our module, which enables them to extract robust and discriminative RF fingerprints. Our approach has been extensively tested through experimental trials, and the results have demonstrated its effectiveness. It is shown that the performance of convolutional neural networks on RF fingerprinting can be improved 1.5% on average, and DAConv-ResNet50 which combined ResNet50 and our Dual Attention Convolution module can achieve 95.6% recognition accuracy on 10 USRP X310. Our source code is available at
https://github.com/zhangweifeng1218/Adaptive_RF_Fingerprinting
.