SADA: semantic-aware perturbation and dual-view alignment for privacy-preserving LLMs

Large language models (LLMs) have attracted widespread attention owing to their strong capabilities in text comprehension and generation. However, despite their popularity in cloud-based inference services, concerns have emerged regarding the potential leakage of sensitive user information. Existing context perturbation methods with differential privacy apply uniform noise to input tokens to mitigate such risks, facing the trade-off between privacy efficiency and generation quality. To narrow this gap, we propose semantic-aware perturbation and dual-view alignment (SADA) for privacy-preserving LLMs inference. SADA introduces two plug-in modules named the perturbation module and alignment module. In the perturbation module, we design ADRTEXT, a semantic-aware local differential privacy (LDP) mechanism that achieves differentiated protection through token-level sensitivity analysis. Then, a novel dual-view alignment strategy is introduced in the alignment module to bridge a randomized perturbed view with an entity-free semantic skeleton, optimizing the trade-off between LDP-induced noise and generation quality. Extensive experiments across multiple datasets demonstrate that SADA achieves effective privacy protection while maintaining a higher generation quality.