DFDK-MSNet: A Deep Fusion Framework of Domain Knowledge and Multiscale 1DCNN for Specific Emitter Identification

More and more devices access the wireless network with the development of wireless communication, making device security a widespread concern. Specific emitter identification (SEI) is significant for the authorization and security of wireless devices and instrumentation. Aiming at improving the identification performance for a large number of wireless devices, a novel deep fusion framework for SEI, named DFDK-MSNet, is proposed. It organically integrates domain knowledge (DK) and multiscale deep networks. The DK is deeply injected into the DFDK-MSNet at the stages of signal decomposition and radio frequency (RF) fingerprint characterization. It is evaluated whether the signal decomposition results are most conducive to individual difference characterization based on the DK of RF mechanism. A multiscale 1DCNN is proposed to extract more subtle and comprehensive features further at a multi and deep level, with the texture RF fingerprint of DK as the input. Massive experiments on a real-world dataset are conducted to verify the effectiveness and superiority of the proposed method, which can greatly increase the number of devices to be identified while ensuring high accuracy. It achieves an identification accuracy above 88% for identifying 60 transmitters even belonging to the same model and batch. The proposed method improves the accuracy by 8% for 60 emitters compared to the comparative methods, and the convergence speed is significantly accelerated with the stable convergence epoch being only 1/6 of the comparative methods. It is significant for the practical authorization scenarios of wireless devices.