Let's Hide from LLMs: An Adaptive Contextual Privacy Preservation Method for Time Series Data
Ubaid Ur Rehman, Musarrat Hussain, Tri D.T. Nguyen, Sungyoung Lee · 2023
The emergence of the Internet of Things (IoT) has evolved various application areas, such as healthcare, smart energy management, and autonomous vehicles. These devices continuously transmit time-series data that can be utilized by a variety of applications to provide personalized services. Recently, Large Language Models (LLMs) have been widely adopted in these application areas to input time-series data into prompts for in-context learning and to retrieve relevant responses accordingly. The time-series data contains sensitive information, and its processing can lead to privacy concerns. Several solutions have been proposed in the literature using differential privacy, which protects single data points or batch-wise privacy preservation through manual configuration of the privacy parameter (ɛ). In this paper, we propose an adaptive contextual privacy preservation method that analyzes the data attributes required for specific application services, acting as context. It then identifies sensitive attributes and adaptively selects the value of ɛ for each data attribute to maintain a balance between privacy and service requirements. The proposed approach was evaluated using power consumption and solar power generation datasets. The results show that the proposed approach dynamically selects the privacy parameter for each data attribute. Moreover, the original and anonymized data were fed into the prompt to assess the textual responses generated by LLM. The results show that our proposed approach achieved an average degree of semantic similarity score of 94.5% for power consumption data and 95.23% for solar power generation data.