GP-DGECN: Geometric Prior Dynamic Group Equivariant Convolutional Networks for Specific Emitter Identification

Abstract

With the rapid development of mobile Internet technology, the number of access devices is increasing exponentially. However, due to inadequate encryption measures or low encryption strength of some devices, illegal access and the easy acquisition of legitimate device user information can lead to privacy breaches and property loss. Recently, physical layer security authentication technology has been adopted to improve the accuracy of identifying illegal devices. However, during signal propagation, noise and channel effects often degrade identification performance. To address this, this paper proposes a dynamic group-equivariant convolutional network based on geometric priors, termed GP-DGECN. This framework combines group-equivariant convolutional layers and dynamic convolution kernel strategies to resolve the limitation of traditional CNN models that only possess translational equivariance. By fully extracting intrinsic signal features, it enhances resistance to high noise and channel effects. Performance tests on real WiFi datasets demonstrate that the proposed framework can achieve an accuracy of 80% under both LoS and NLoS channel scenarios. Even under strong interference from SUIA channel parameters, it can achieve a recognition accuracy of nearly 60%.