همايش ملی پيشرفت های معماری سازمانی

صفحه اصلی / نهمین همایش پیشرفت‌های معماری سازمانی ایران

Privacy-Preserving LLM Inference in Intelligent Enterprises: Threats, Defenses, and Research Frontiers

نویسندگان :

Reza Fotohi¹ Fereidoon Shams Alie²

1- دانشگاه شهید بهشتی 2- دانشگاه شهید بهشتی

کلمات کلیدی :

Large Language Models (LLMs)،Privacy-Preserving Inference،Membership Inference Attacks،Model Reconstruction Attacks،Intelligent Enterprises

چکیده :

Large Language Models (LLMs) are increasingly integrated into intelligent enterprises to support critical services across healthcare, finance, and government domains. While their inference capabilities offer unprecedented utility, they also introduce significant privacy risks, as both user queries and model outputs are directly exposed during operation. Recent studies have demonstrated the vulnerability of LLMs to inference-time attacks, including Membership Inference Attacks (MIA) and Model Reconstruction Attacks (MRA), which can extract or reconstruct sensitive enterprise data. Although multiple defense mechanisms have been proposed, they often suffer from limited scope, high computational overhead, or reduced utility, making them unsuitable for enterprise-scale deployment. This survey provides a structured and comprehensive overview of privacy-preserving LLM inference. We first review the main threats in the inference stage and categorize existing defense mechanisms into input-level defenses, output-level defenses, and adaptive/feedback-based strategies. Each category is examined for its strengths, weaknesses, and applicability in real-world enterprise settings. We further analyze key research gaps, including the lack of dual-attack robustness, limited enterprise-grade evaluations, and the absence of standardized benchmarks. Building on these insights, we highlight open challenges and outline promising future directions. By consolidating the state of the art, this work serves as a reference for both academia and industry, guiding the development of secure, efficient, and compliant LLM deployment in intelligent enterprises.