Guardians of Privacy: Leveraging LLMs in Assistive Robotic Systems for Healthcare

This paper presents a Privacy-Aware Assistive System (PAAS) designed to mitigate privacy risks associated with integrating cloud-based Large Language Models (LLMs) into resource-constrained systems like Socially Assistive Robots (SARs), in healthcare settings, which results in seeking cloud-based solutions. While cloud-based LLMs substantially enhance patient care through more natural and effective interactions, their use raises significant privacy concerns due to the potential exposure of sensitive healthcare data to external services. To address this challenge, PAAS employs the Principle of Least Privilege (PoLP), by fine-tuning domain-specific LLMs to accurately identify user intents and extract necessary parameters used to query a structured database from unstructured natural language inputs without exposing sensitive data directly to third-party services. The paper introduces an algorithm for generating domain-specific datasets, facilitating precise intent classification and custom entity recognition essential for querying internal databases securely. Performance comparisons among fine-tuned models were conducted using varying complexities of user requests, including basic, context-sensitive, and ambiguous interactions. Results demonstrate robust performance, with the GPT-4O-mini fine-tuned model achieving an F1 score of up to 95% across multiple tests conducted at different times and days. PAAS effectively facilitates high-quality, natural user interactions through the advanced capabilities of LLMs while rigorously maintaining user privacy. Future research will address improving the system's resilience to linguistic ambiguities and further advancing its privacy safeguards.