LLMs as Linguistic Chameleons: Decoupling Semantics and Structure for Privacy-Preserving Communication

Large Language Model (LLM) APIs are increasingly embedded in privacy-sensitive workflows, yet ensuring inference-time privacy without sacrificing task utility remains an open challenge. Existing privacy-preserving approaches typically retain most of the original semantic content to maintain downstream performance. However, this retention leaves exploitable cues that adversaries can use to reconstruct the original text. The authors, Yuzhu Mao and Liang Zhao, investigate **semantic decoupling**: replacing original semantics with alternative content while preserving the structural properties required for LLM reasoning. This motivates **CROSS-MAP**, a bidirectional framework that maps private inputs into a different semantic domain before inference and recovers the corresponding outputs afterward. The core tension is between semantic divergence (to defeat reconstruction) and semantic consistency (to preserve utility).

Experiments evaluate CROSS-MAP across multiple attack settings designed to reconstruct the original input from the transformed representation. The authors report that CROSS-MAP **reduces reconstruction success** compared to existing baselines, indicating stronger privacy protection. At the same time, it **outperforms existing baselines in utility**, meaning the recovered outputs remain useful for downstream tasks. The results suggest that semantic decoupling can achieve a better privacy-utility trade-off than methods that preserve most of the original semantics. The paper does not provide specific numerical values in the abstract; detailed metrics are expected in the full text.