To address the issue of information leakage caused by key theft during transmission, a symmetric encryption end-to-end communication system based on channel state information (CSI) was proposed. The proposed system employed convolutional neural networks to construct the transmitter, receiver, and key generator, optimizing the encoding and decoding process in an end-to-end manner. At the same time, it leveraged the reciprocity, random time-variability, and spatial uniqueness of wireless channels to measure the CSI and generate keys from legitimate users, encrypting the original information. The simulation results demonstrate that under Rayleigh fading, Rician fading, and frequency-selective multipath fading channels, the bit error rates (BER) of the proposed system is lower than that of baseline models such as the symmetric encryption system based on deep convolutional generative adversarial networks within the tested signal-to-noise ratio (SNR) range. In high-SNR scenarios, the improvement of BER can reach 18 dB. Additionally, in four attack scenarios, such as brute force and key leakage attacks, the BER of eavesdropper is approximately 0.5, indicating an inability to decrypt the information. The proposed system eliminates the need for key distribution, reducing the BER while enhancing eavesdropping resistance, thus providing a novel approach for secure communication.