In the era of big data, the availability of massive amounts of information make privacy protection more necessary than ever. Among a variety of anonymization mechanisms, microaggregation is a common approach to satisfy the popular requirement of k-anonymity in statistical databases. In essence, k-anonymous microaggregation aggregates quasi-identifiers to hide the identity of each data subject within a group of other k – 1 subjects. As any perturbative mechanism, however, anonymization comes at the cost of some information loss that may hinder the ulterior purpose of the released data, which very often is building machine-learning models for macrotrends analysis. To assess the impact of microaggregation on the utility of the anonymized data, it is necessary to evaluate the resulting accuracy of said models. In this work, we address the problem of measuring the effect of k-anonymous microaggregation on the empirical utility of microdata. We quantify utility accordingly as the accuracy of classification models learned from microaggregated data, evaluated over original test data. Our experiments indicate, with some consistency, that the impact of the de facto microaggregation standard (Maximum Distance to Average Vector, MDAV) on the performance of machine-learning algorithms is often minor to negligible for a wide range of k, for a variety of classification algorithms and data sets. Furthermore, experimental evidences suggest that the traditional measure of distortion in the community of microdata anonymization may be inappropriate for evaluating the utility of microaggregated data.

Hierarchical text classification consists of 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 it 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.

Our behavior —the way we talk, walk, act, or think— is unique and can be used as a biometric trait. It also correlates with sensitive attributes like emotions and health conditions. With more and more behavior tracking techniques (e.g. fitness trackers, mixed reality) entering our everyday lives more of our behavior is captured and processed. Hence, techniques to protect individuals’ privacy against unwanted inferences are required, before such data is processed. To consolidate knowledge in this area, we are the first to systematically review suggested anonymization techniques for behavioral biometric data. We taxonomize and compare existing solutions regarding privacy goals, conceptual operation, advantages, and limitations. Our categorization allows for the comparison of anonymization techniques across different behavioral biometric traits. We review anonymization techniques for the behavioral biometric traits of voice, gait, hand motions, eye gaze, heartbeat (ECG), and brain activity (EEG). Our analysis shows that some behavioral traits (e.g., voice) have received much attention, while others (e.g., eye gaze, brain activity) are mostly neglected. We also find that the evaluation methodology of behavioral anonymization techniques can be further improved.

Continuous data publishing consists in the republication of updating microdata. The most relevant syntactic notions in continuous data publishing are based on m-invariance. This notion enforces that no user can be distinguished among, at least, m - 1 other users, each with distinct secret data. To achieve m-invariance, the existing methods must first alter the dataset to satisfy a property called m-eligibility. Essentially, a dataset can be made m-invariant if and only if it satisfies the m-eligibility constraint. Although guaranteeing the m-eligibility property is a crucial step, no theoretical study of the best strategies to achieve it has been carried out. This paper performs such a study by giving strategies and demonstrating their optimality under two approaches: insertion of counterfeit tuples and partial publication. The empirical evaluation of our proposal shows a significant reduction on the number of modifications needed to enforce m-eligbility of up to 41% with respect to the literature.