A New Utility Evaluation Framework for Data Anonymization in the Context of Mobility

Sharing urban mobility and public transportation data is critical
to use the mobility infrastructure of cities to its fullest potential. For
data protection reasons, however, the disclosure of data to the public
is restricted and only permitted if the anonymity of each individual
associated with the dataset can be guaranteed. To achieve anonymity
in a given dataset, numerous approaches can be applied, while each ap-
proach follows a dierent denition of anonymity. One of the most used
denitions is k-anonymity, which builds on the building of equivalence
classes so that each row in a dataset belongs to an equivalence class
that contains at least k rows that cannot be distinguished. Naturally,
this can be achieved by multiple realizations. However, the question is
which realization will provide the highest utility for future real-world
applications. Currently, abstract metrics are used to assess the utility
of dierent k-anonymizations, based on the structure of the dataset.
However, these abstract metrics do not properly reect the usefulness
of the anonymized datasets in real-world applications. Hence, in this
work, we provide a novel framework that helps to evaluate the given
abstract metrics from the literature in terms of their performance in
measuring utility in the context of urban mobility. To do this, we de-
ne a set of potential data science use cases that can be derived from
a publicly available dataset on taxi drives and compute multiple real-
izations of k-anonymity. By training prediction models on the original
dataset and the anonymized datasets and comparing the corresponding
performance decrease with the abstract metrics from the literature, we
are able to derive recommendations on the usage of abstract metrics
to evaluate the utility of potential realizations to achieve k-anonymity.