The interaction and fusion of cross-domain features and feature semantics are crucial for obtaining cross-domain generalized feature representations in Network Intrusion Detection (NID) tasks. Existing interaction methods typically treat features within a single domain as independent entities, utilizing various feature engineering techniques for intra-domain feature selection, interaction, and fusion. Although these methods perform well in domain-specific NID tasks, their cross-domain generalization capabilities remain limited. To address these challenges, this paper proposes a novel Cross-Domain Semantic Fusion (CDSF) framework, which integrates cross-domain feature semantic matching, multi-task mapping, and multi-granularity feature fusion into a unified process. This framework ensures effective transformation and fusion of discrete and continuous features across domains. Comprehensive experiments on multiple benchmark NID datasets demonstrate that the proposed method achieves competitive performance in terms of both accuracy and cross-domain generalization.