Privacy Preservation in Textual Data: A Systematic Mapping Study on Differential Privacy and Semantic Similarity
Abstrak
Background: AI and machine learning increasingly depend on large-scale textual data, which often embeds personally identifiable and sensitive information. In this context, privacy-preserving processing of unstructured text has become essential to mitigate disclosure and re-identification risks, especially in pipelines that rely on semantic representations and similarity measures. Goal: This study aims to map state-of-the-art techniques for privacy-preserving textual data analysis and to characterize how privacy mechanisms relate to semantic similarity and the identification of infrequent (rare) textual patterns. Method: We conducted a Systematic Mapping Study (SMS) following established guidelines, retrieving peer-reviewed publications from four major digital libraries (2010–2025). The selected studies were screened using inclusion/exclusion criteria and a quality checklist, and the extracted data were synthesized through frequency-based and thematic mapping aligned with three research questions: (i) privacy-preserving techniques used in textual data analysis, (ii) computational approaches (data science and language-model-based methods) supporting such mechanisms, and (iii) techniques adopted for semantic similarity and rare-event oriented text analysis under privacy constraints. Results: The mapping shows that anonymization/de-identification, differential privacy, and federated learning are among the most recurrent privacy-preserving approaches reported for text. It also highlights the prevalence of NLP pipelines and transformer-based models (e.g., BERT variants and large language models) as supporting components, typically combined with classic semantic similarity techniques such as vector-space representations, embeddings, topic modeling, and cosine similarity. Rare-event detection appears less frequently, suggesting an emerging gap and opportunities for future research on privacy-aware pipelines for low-frequency phenomena in text. The results provide an evidence-based overview to support the design of guidelines and best practices for privacy-preserving text analytics in software intensive settings.
Identifier
- DOI: https://doi.org/10.5753/cibse.2026.42448
- OpenAlex: https://openalex.org/W7162788076