AI-Native Network with Clean–Impaired Semantic Preservation for Robust Semantic Communications in Space

Semantic communication offers a method for task-oriented information delivery in artificial intelligence (AI)-enabled space networks, but compact semantic representations remain vulnerable to fading, noise, and structured corruption during transmission. This work proposes Clean–Impaired Semantic Preservation for Robust Semantic Communications (CISP-SemCom), a link-level semantic communication framework that explicitly learns from the clean latent representation and its communication-impaired counterpart generated from the same source sample. In a shared projection space, class-aware Clean–Impaired Contrastive Alignment (CI-Con) preserves cross-sample discrimination while excluding same-class non-matched samples from the negative set, and Semantic Preservation (SP) directly promotes consistency within each matched clean–impaired pair. In a ten-seed controlled same-backbone evaluation on CIFAR-10 under Rician fading, CISP-SemCom increases the normalized area under the robustness curve (AURC) evaluated over the signal-to-noise ratio (SNR) (SNR-AURC) from 66.85 ± 2.36 to 72.48 ± 1.24, corresponding to a paired gain of 5.63 AURC points with a 95% confidence interval of [3.75, 7.51]. Mechanism diagnostics show that SP-only optimization admits projector collapse, and that class-agnostic contrastive learning introduces same-class false-negative pressure. A direct class-aware ablation further indicates that most of the measured task-robustness gain comes from class-aware CI-Con, while SP acts as an explicit matched-pair consistency regularizer. On EuroSAT RGB, SNR-AURC increases from 65.53 ± 1.80 to 75.35 ± 1.58, showing that the robustness improvement is also observed on a second dataset. Decoder-independent latent probes further show improved accessibility of task-relevant class information, while the measured batch-1 inference latency remains similar to Backbone. These results support clean–impaired semantic preservation as an effective strategy for robust task-oriented transmission under satellite-inspired link impairments.