RBDPM: Risk-Based Differential Privacy Model for Trajectory Data

Personal safety applications enable users to communicate emergency situations to relevant third parties and local authorities. Location-Based Services play a crucial role in the capture and exchange of data, including location and personal identifiable information, to better inform emergency response efforts. Maximising the effectiveness of these safety applications requires the data to be accurate and informative yet prevent the exposure of sensitive user information. Current solutions often fail to adequately protect this sensitive data in the attempt to maintain accurate and useful information for emergency response. Therefore, personal safety solution safety applications should be able to protect the privacy of individuals without compromising the overall utility and accuracy of the data. This thesis presents a Risk-Based Differential Privacy Model for Location Data that is designed to assess safety-critical factors and attributes associated with users and scenarios to provide a dynamic balance for trajectory data utility and privacy trade-off. The model assesses the safety-critical factors facing the user from the data and quantifies the risk in the Hazard Assessment Module. The quantified risk informs the level of privacy parameters in the Privacy Preservation Module, which will determine the levels of noise to be added to the dataset in the Noise Application Module to ensure that lower risk levels can afford maximum privacy, whereas high-risk scenarios will result in reduced privacy without losing data utility. The resulting noise-injected trajectory dataset is processed using the Linear Regression model to validate this concept and evaluate the impact of data utility and privacy trade-off in the dataset during processing. The performance of the dataset to retain utility while ensuring privacy during processing is analysed using evaluation criteria metrics that explore the efficiency, generalisation, and robustness of the dataset. The metrics outcome show that the noise-injected dataset can maintain good data utility while safeguarding the privacy of the user when processed. The outcome emphasises the importance of exploring factors and attributes associated with safety-critical data by the user and the dataset to dynamically find the optimal balance for the data utility and privacy trade-off.