Authors - Maricela Pinargote-Ortega, Marely del Rosario Cruz Felipe, Carlos Manuel Lucas Aragundi, Iter Alexander Posligua Solorzano Abstract - Open data is often associated with objectives linked to fostering innovation and economic growth, political accountability and democratic participation, and public sector efficiency. However, data privacy has been frequently cited as a challenge for open data publication and processing. This paper uses a 9780-row dataset from the 2025 community engagement survey of the Philippine National Police Regional Office 5 to synthesize a privacy-preserving dataset using natural language processing and the Laplace mechanism with a total Privacy Loss Budget (PLB) value of 1. The text dataset fields with the highest privacy risk were replaced with generated topic models and corresponding overall sentiment values. The dataset fields were then categorized into four blocks, grouping variables that require correlations to be preserved. Noise was added to the four blocks using the Laplace mechanism, generating a privacy-preserved synthetic query robust to de-anonymization attacks. The synthesized dataset shows minimal distortion from the original dataset, with mean shifts of less than 0.25, and preserving key variable correlations, while significantly increasing data subject privacy. End-user validation confirmed that the synthetic dataset is suitable for both data sharing and joint processing without sacrificing the accuracy of analysis results. This study demonstrated that a differentially private synthetic data generation pipeline combining natural language processing and the Laplace mechanism (ε = 1) can substantially enhance data subject privacy while preserving the analytical utility of a real-world public sector survey dataset.
Open Zoom