Deep Learning‐Driven Semantic Communication With Attention Modules

ABSTRACT In this study, an innovative architecture is proposed to enhance the performance of semantic communication networks by leveraging deep learning and joint source‐channel coding. A fundamental challenge in this field is the strong dependence of conventional networks on a fixed signal‐to‐noise ratio (SNR) during training, which leads to performance degradation under varying channel conditions. To address this limitation, we introduce a novel attention‐based approach that enables dynamic adaptation to different SNR levels, ensuring more stable and optimized communication performance. The proposed model learns more generalized features that exhibit greater resilience to channel variations. To evaluate its effectiveness, extensive simulations were conducted, comparing the performance of the proposed architecture with DeepSC, a state‐of‐the‐art benchmark model in the field. While the baseline model, trained at a single SNR, experiences performance drops under mismatched conditions, the proposed model, trained across a range of SNRs, achieves improvement of 16.2%, 30.8%, 42.8%, and 53.8% for 1, 2, 3, and 4‐gram precisions, respectively, in bilingual evaluation understudy score and an 11.4% increase in sentence similarity across challenging low‐SNR conditions. Furthermore, the model maintains robust performance with 48% less training data, highlighting its efficiency and data efficiency under practical constraints. These gains confirm the model's superior adaptability and high‐quality data reconstruction under diverse conditions. The results of this study underscore the significant benefits of attention‐based architectures in semantic communication, particularly in environments with unpredictable channel variations, and highlight their potential for reliable deployment in real‐world applications.