Privacy risk assessment method incorporating sensitivity and correlation with empirical study

Document Type

Article

Publication Date

1-1-2026

Abstract

Introduction. User-generated content (UGC) has emerged as a prominent vector for privacy breaches, especially due to the context-dependence of data sensitivity and vulnerabilities introduced by data correlations. These challenges highlight the growing limitations of traditional assessment method s. Method. This study proposes a privacy risk quantification method integrating both attribute sensitivity and inter-attribute association, with an experimental validation conducted on the ‘Friend Identification ’ section of the https://muchong.com . A BERT-BiLSTM-CRF deep learning model is utilized for the automatic identification of attributes from unstructured text. Using a predefined privacy data lexicon, attribute sensitivity is quantified, and pointwise mutual information (PMI ) is introduced to measure attribute associations. Combined with a privacy subject identification factor, these elements collectively quantify privacy ris k values, followed by risk level classification. Results. Ablation experiments and manual validation have confirmed the feasibility of the proposed scheme, demonstrating its capability to identify, assess, and classify privacy risks in unstructured textual data with broad applicability. Conclusion. The study validates the proposed solution theoretically, technically, and empirically, overcoming the limitations of traditional isolated-field evaluation paradigms. The method can be extended to high-sensitivity domains such as healthcare and finance, providing a basis for dynamic, risk-informed classification policies.

Publication Title

Information Research

DOI

10.47989/ir31iConf64277

Volume

31

Issue

iConf (2026)

First Page

664

Last Page

674

Share

COinS