Abstract: The rapid development of wireless communication technology has brought unprecedented challenges to the communication security of wireless communication systems. Effectively identifying and analyzing the physical layer signals of underwater acoustic communication under non-intervention conditions is becoming increasingly important. A method for individual identification of radiation sources in underwater acoustic communication signals based on cyclic spectral features is proposed in this paper. Different roll-off factors of root raised cosine filters are used to characterize individual underwater acoustic communication signal radiation sources. A lightweight neural network model, MobilenetV3 small, suitable for underwater acoustic communication signals, is designed. The cyclic spectrum is used as the network input to achieve specific emitter identification of 5 binary phase shift keying (BPSK) modulated radiated sound source signals. Simulation results demonstrate that the proposed method outperforms the traditional convolutional neural network VGG16 in terms of running speed, parameter quantity, and loss rate, thus proving the effectiveness of the algorithm.