Trajectory analysis holds many promises, from improvements in traffic management to routing advice or infrastructure development. However, learning users' paths is extremely privacy-invasive. Therefore, there is a necessity to protect trajectories such that we preserve the global properties, useful for analysis, while specific and private information of individuals remains inaccessible. Trajectories, however, are difficult to protect, since they are sequential, highly dimensional, correlated, bound to geophysical restrictions, and easily mapped to semantic points of interest. This paper aims to establish a systematic framework on protective masking and synthetic-generation measures for trajectory databases with syntactic and differentially private (DP) guarantees, including also utility properties, derived from ideas and limitations of existing proposals. To reach this goal, we systematize the utility metrics used throughout the literature, deeply analyze the DP granularity notions, explore and elaborate on the state of the art on privacy-enhancing mechanisms and their problems, and expose the main limitations of DP notions in the context of trajectories.
Hierarchical text classification consists in classifying text documents into a hierarchy of classes and sub-classes. Although artificial neural networks have proved useful to perform this task, unfortunately they can leak training data information to adversaries due to training data memorization. Using differential privacy during model training can mitigate leakage attacks against trained models, enabling the models to be shared safely at the cost of reduced model accuracy. This work investigates the privacy-utility trade-off in hierarchical text classification with differential privacy guarantees, and identifies neural network architectures that offer superior trade-offs. To this end, we use a white-box membership inference attack to empirically assess the information leakage of three widely used neural network architectures. We show that large differential privacy parameters already suffice to completely mitigate membership inference attacks, thus resulting only in a moderate decrease in model utility. More specifically, for large datasets with long texts we observed Transformer-based models to achieve an overall favorable privacy-utility trade-off, while for smaller datasets with shorter texts convolutional neural networks are preferable.