Transformer-Based Shared Embedding for Multiple Access in Semantic Communications

초록

This paper proposes a Transformer-based shared embedding (SE) scheme as an artificial intelligence (AI)-native multiple access for semantic communications. In SE, multiple users jointly occupy a high-dimensional embedding space and are separated by learnable user-specific positional masks together with attention-driven demultiplexing, so that inter-user interference is suppressed directly in the embedding domain. The resulting interference structure closely resembles interference alignment (IA), wherein residual multiuser energy is compressed into a low-rank subspace that the decoder can effectively attenuate. We develop closed-form expressions for signal power, effective signal-to-interference-plus-noise ratio (SINR), symbol error rate (SER), and per-user capacity, explicitly characterizing the roles of intra-user coherence and an alignment coefficient that quantifies attention-induced interference suppression. These formulas yield theoretical baselines that situate SE performance between a no-alignment lower bound and an ideal non-orthogonal multiple access (NOMA)-inspired upper bound and explain capacity gains in dense regimes via statistical multiplexing and alignment. Experimental evaluations are carried out under Rayleigh fading channels, and the results show that SE consistently outperforms the dedicated embedding (DE) scheme, which assigns each user an exclusive portion of the embedding space, across different channel conditions and user densities. The empirical SE curves match the theoretical predictions with fitted alignment factors and approach the NOMA-inspired bound as the embedding dimension increases, thereby showing that the embedding space can function as a new axis of multiple access complementary to time, frequency, and code domains. Finally, we outline adaptive and generalized training strategies that extend beyond symbol-level recovery to semantic-level classification and reconstruction tasks for next generation networks.

키워드