Striking the Balance: Generalization vs. Memorization in Anonymization and De-anonymization through LLMs

Text anonymization aims to enable the secure sharing of information between parties. One of the main challenges in data anonymization is achieving a balance between ensuring data privacy and maintaining data utility. To address these challenges, recent studies have explored the use of Large Language Models (LLMs), which have shown improved performance on datasets from Europe. Based on these findings, this paper aims to create a dataset from less explored parts of the world, specifically Africa, to assess the relevance of LLMs on diverse datasets and to discuss the generalization of the results. Additionally, this paper proposes an evaluation framework for assessing various anonymization techniques, including those utilizing LLMs. The performance of these techniques is assessed using several metrics, such as BERTScore for semantic evaluation and Information Loss for utility preservation.