An Investigation of Power Amplifier Feature for Deep Learning Based RF Fingerprint Identification

Radio frequency fingerprint identification (RFFI) employs unique hardware impairments for device authentication. Power amplifier (PA) nonlinearity is an important hardware impairment feature, which has been widely used for RFFI. In this paper, we modeled and estimated PA parameters and studied their effects on RFFI. In particular, a memory polynomial model was used to capture the detailed features of PA, involving polynomial order, memory depth and coefficient matrix. These parameters are twisted and we designed grid search and deep learning-based algorithms to find optimal parameters, with a tradeoff of computational complexity and RFFI classification accuracy. We created a testbed consisting of 60 IEEE 802.15.4 devices and a universal software radio peripheral (USRP) X310 software-defined radio (SDR) platform as the receiver. The PA function of our devices can be disabled when needed, which allowed us to generate signals with and without PA features. Experimental results demonstrated that our proposed algorithms can estimate the PA parameters. Simulated signals with the estimated PA parameters were also demonstrated to retain the PA features accurately.