2026-05-18T10:59:25Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-2014612025-05-12T19:35:22Zoai_pmh:alloai_pmh:repo_items

201461 https://hdl.handle.net/2286/R.2.N.201461 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2025 73 pages Masters Thesis Academic theses en Choudhary, Sumeet Yau, Stephen S Yang, Yingzhen Baek, Jaejong Arizona State University Partial requirement for: M.S., Arizona State University, 2025 Field of study: Computer Science Protecting user’s privacy while doing text-based interactions with Large Language Models (LLMs) has become an important concern, with users disclosing sensitive data pertaining to personally identifiable information (PII), financial data, and health data when providing a prompt. Given the intelligence of the most recent LLMs, they operate and achieve better results when they learn as much as possible about the user, thereby learning and absorbing all the data provided by the user. Although a couple of processes are in place to mitigate the risk of sensitive information shared with these models, none effectively address the significant challenge of protecting user data in real time when engaging in user-LLM interactions. In this process, a Dynamic Privacy Preservation approach is presented to dynamically analyze, classify, and intervene to protect sensitive data within user inputs and surrender them in real time. The approach utilizes a hybrid sensitivity detection approach, including Named Entity Recognition (NER), regex matching, and context-based scoring to detect sensitive data accurately. In the solution, adaptive privacy mechanisms are implemented based on the sensitivity level of the keyword assigned by the algorithm, such as redaction, anonymization, and generalization to provide a balance between protecting user/private data and maximizing data utility. The evaluation results demonstrate that the presented approach performs substantially better than existing methods. In conclusion, the approach fills an important gap that, if widely adopted can be beneficial towards the multiple businesses that are currently not leveraging LLM services to avoid any organization/user data leakage. Computer Science An Approach to Dynamic Privacy Preservation for User Interactions with Large Language Models