Proceedings of the
The Nineteenth International Conference on Computational Intelligence and Security (CIS 2023)
December 1 – 4, 2023, Haikou, China
RF Fingerprinting Identification with Low-Power Neural Networks
School of Cyberspace Security, Hainan University, China.
ABSTRACT
Deep learning methods have generated significant results in the field of radio frequency fingerprint identification.
But many existing approaches have high requirements on computational resources, which is difficult to save cost.
In order to make it more widely applicable across various fields, we focus on researching how to implement this
technology in resource-constrained environments. We analyzed layerwise compression and channel compression
strategies on neural networks, and used channel compression as the main scheme. The method has a wider range of
application scenarios, while saving computational resources more significantly.We chose three specific compression
ratios:3/4, 1/2, and 1/4. We applied this compression technique to the well-performing deep residual networks
(DRSN) model and used the method of knowledge distillation to compensate for the accuracy loss caused by
compression. Compared with the existing methods, our method show the state-of-the-art performance even after
substantial compression of the student model, fully demonstrating the effectiveness of our experimental method.
Our experiments show that the channel compression strategy used on the DRSN model can achieve a compression
rate of 90.9%. The maximum accuracy can reach 93.22% when the SNR is 10dB.
Keywords: Wireless Security, Radio frequency fingerprinting, deep learning, model compression, knowledge distillation.
Download PDF
School of Cyberspace Security, Hainan University, China.
ABSTRACT
Deep learning methods have generated significant results in the field of radio frequency fingerprint identification. But many existing approaches have high requirements on computational resources, which is difficult to save cost. In order to make it more widely applicable across various fields, we focus on researching how to implement this technology in resource-constrained environments. We analyzed layerwise compression and channel compression strategies on neural networks, and used channel compression as the main scheme. The method has a wider range of application scenarios, while saving computational resources more significantly.We chose three specific compression ratios:3/4, 1/2, and 1/4. We applied this compression technique to the well-performing deep residual networks (DRSN) model and used the method of knowledge distillation to compensate for the accuracy loss caused by compression. Compared with the existing methods, our method show the state-of-the-art performance even after substantial compression of the student model, fully demonstrating the effectiveness of our experimental method. Our experiments show that the channel compression strategy used on the DRSN model can achieve a compression rate of 90.9%. The maximum accuracy can reach 93.22% when the SNR is 10dB.
Keywords: Wireless Security, Radio frequency fingerprinting, deep learning, model compression, knowledge distillation.
Download PDF