In the use case on text anonymisation of sensitive data (Use Case 4), led by the Universidad de León, EOSC-SIESTA focuses on enabling the safe use of textual data for machine learning applications where real datasets cannot be shared due to their sensitive nature.
To address this limitation, a synthetic dataset of text documents simulating cyber-incident reports is being developed, as real data is not available for privacy reasons. This dataset is carefully annotated so that (a) personal data is identified and categorised into specific types, and (b) each document is labelled according to its category.
In parallel, a solution for data anonymisation and pseudonymisation is being developed, based on the identification of sensitive information in text through Named Entity Recognition (NER) techniques.
While current approaches typically rely on masking, suppression, pseudonymisation or noising of sensitive data, the proposed approach combines structured annotation, NER-based detection, and synthetic data generation. This allows replacing original sensitive content with realistic synthetic data, preventing disclosure while preserving the utility of the dataset. As a result, machine learning models can be trained to classify text into the same categories with comparable performance to that achieved using the original data