Swin-SemantIC: A Transformer-Driven Enhancement Scheme for Semantic Interference Cancellation

This paper proposes an enhanced semantic interference cancellation (SemantIC) technique, Swin-SemantIC, which improves the efficiency of cross-domain noise mitigation and anti-interference performance by optimizing the semantic auto-encoder in the original SemantIC system. Grounded in the Wyner-Ziv theorem, the original SemantIC system constructs a Turbo loop by cascading a channel decoder and a CNN-based semantic auto-encoder, enabling alternating denoising in the signal and semantic domains to enhance received signal quality. However, the inherent limitation of CNNs in local feature extraction constrains their capability to capture global semantic content of information, thereby degrading the quality of side information. To address this, Swin-SemantIC employs a Swin Transformer-based semantic auto-encoder to strengthen global dependency capture, enabling more efficient and reliable side information transmission within the Turbo loop. Simulation results demonstrate that Swin-SemantIC achieves further performance gains, with particularly significant advantages under low signal-to-noise ratio conditions, meanwhile incurring no additional channel resource consumption.