Design of End-to-End Communication System Under CBAM-CGAN Channel

To address the limitations of generation accuracy and training instability inherent in conditional generative adversarial network (CGAN)-based methods, an enhanced end-to-end (E2E) communication system was proposed in the present study. The proposed framework incorporates channel modeling via a convolutional block attention module (CBAM), and exploits the adaptive data generation capabilities of CGANs alongside the CBAM’s proficiency in extracting multidimensional spatial and channel-dependent features, thereby enabling accurate representation of unknown communication channels. To mitigate the training instability typically associated with CGANs, a maximum mean discrepancy (MMD) loss component was introduced into the generator’s objective function, facilitating improved convergence behavior. Experimental evaluations demonstrate that the proposed model achieved superior performance relative to existing GAN-based approaches, particularly in reducing bit error rate (BER) across a range of signal-to-noise ratio (SNR) conditions, including additive white Gaussian noise (AWGN) channels, Rayleigh fading environments, frequency-selective multipath scenarios and optical fiber channels. Notably, at high SNR regimes, the system yielded a BER improvement of up to 2 dB. These findings substantiate the proposed method’s robustness across diverse channel conditions and highlight its potential as a novel framework for enhancing the performance of communication systems.