Online Social Networks (OSNs) have become ubiquitous in the activities of people recently. However, a large number of disclosing private information are posted by online social network users unconsciously every day, and some users may face undesirable consequences, e.g., identity theft. Consequently, the significance of privacy information detection for users of OSNs turns out to be important. A large number of studies have been dedicated to corporate privacy leakage analysis. Whereas, there are very few studies that detect privacy revealing for individual OSNs users.

With this motivation, this thesis aims to propose an automated privacy information detection approach to effectively detect and classify privacy revealing information for individual users. It comprises two steps: detecting privacy information leaks and classifying them into fine-grained categories. In the first step, a deep-learning based model is built to recognise privacy-related entities in a real-world data set, which has achieved a considerable performance based on the experimental results and case studies. In the second step, a semantic phrase similarity degree approach is developed to automatically classify privacy-related entities into fine-grained privacy entities based on a built privacy domain ontology. Finally,extensive experiments are conducted to validate the proposed privacy information approach, and the empirical results demonstrate its superiority in assisting OSNs’s users to avoid the privacy leakage.

This work provided a complete approach to handle privacy information detection on online social networks, which is essential for individuals to mitigate their privacy leakage.