Text-Based Personas for Simulating Privacy User Studies

Traditional privacy user studies are often hindered by high costs and limited scalability. We introduce PersonaSimulator, a framework that simulates these studies using LLMs. By grounding synthetic personas in real-world survey data, PersonaSimulator creates representative individuals that mirror target populations in terms of demographics and attitudes. The system employs theoretical templates—such as Privacy Calculus and Protection Motivation Theory—to standardize decision-making logic. Through an iterative optimization procedure, these personas are refined to be concise and predictive of unseen human responses. Evaluated across five datasets, PersonaSimulator accurately models individual and population-level privacy preferences, providing a cost-effective tool for pilot studies and survey iteration.

• Anti abuse