Memory-Aware Privacy Protection for Large Language Models in Education
Abstract
The integration of Large Language Models (LLMs) in education has revolutionized. However, these models pose significant privacy risks, including data memorization and unintended exposure of sensitive student information. Ensuring data security and regulatory compliance while maintaining LLMs' educational benefits is a critical challenge. This study evaluates ChatGPT-4o's memorization tendencies and privacy risks in educational applications, analyzing their potential for data leakage through adversarial querying. We propose privacy-preserving techniques, including data anonymization ( -anonymity, -diversity) and prompt tuning, to mitigate these risks while preserving student performance assessment capabilities. Empirical experiments on real-world student datasets demonstrate that data anonymization reduces privacy leakage by 74.50%, while optimized prompt tuning lowers personally identifiable information exposure by 57%. Additionally, the combined approach of anonymization and prompt tuning keeps the privacy protection score (PPS) at 7.5%. Although stronger privacy introduces slight trade-offs in data utility, the root mean squared error and mean absolute error increase by 3.44 on average, the techniques maintain a balance between accuracy and security. These findings offer actionable insights for developing secure, ethical, and regulation-compliant AI-powered educational technologies that enhance learning outcomes while safeguarding privacy. Therefore, a critical challenge in the current research landscape is harnessing the benefits of LLMs in the education domain while effectively mitigating privacy risks. This study aims to explore the most advanced GPT-series LLMs, ChatGPT-4o, research the advantages of analyzing student performance, and examine the associated privacy concerns. We first explore the application of ChatGPT-4o in education, focusing on its capability to evaluate student performance. Subsequently, we analyze potential privacy risks that arise during data collection and the extent to which ChatGPT-4o retains and potentially exposes sensitive information, assessing its memorization tendencies and associated privacy risks. To mitigate these risks,