Attention-Driven Interference Alignment for Multi-User Semantic Communications

초록

This paper presents an attention-driven interference alignment (IA) scheme for multi-user semantic communications that operates in the semantic embedding domain, unlike conventional IA in the baseband domain. To realize this, user-specific masking and attention-driven decoding are employed over a shared embedding space, which is mathematically modeled to characterize attention-based IA in the embedding domain and to enable implicit user separation without explicit orthogonalization. Built on a conventional cross-entropy (CE) loss, the proposed orthogonalization regularization further reinforces interference alignment by aligning user-specific semantic subspaces and suppressing inter-user interference. Experiments under Rayleigh fading channels show that the proposed approach outperforms both a resource-partitioned baseline with separate embeddings and a CE-only loss model without orthogonalization regularization, showing the effectiveness of an attention-driven IA solution for multi-user semantic communication.

키워드