A widely used method to ensure privacy of unstructured text data is the multidimensional Laplace mechanism for dX-privacy, which is a relaxation of differential privacy for metric spaces. We identify an intriguing peculiarity of this mechanism. When applied on a word-by-word basis, the mechanism either outputs the original word, or completely dissimilar words, and very rarely outputs semantically similar words. We investigate this observation in detail, and tie it to the fact that the distance of the nearest neighbor of a word in any word embedding model (which are high-dimensional) is much larger than the relative difference in distances to any of its two consecutive neighbors. We also show that the dot product of the multidimensional Laplace noise vector with any word embedding plays a crucial role in designating the nearest neighbor. We derive the distribution, moments and tail bounds of this dot product. We further propose a fix as a post-processing step, which satisfactorily removes the above-mentioned issue.

International audience

Personal Data Management Systems (PDMSs) arrive at a rapid pace providing individuals with appropriate tools to collect, manage and share their personal data. At the same time, the emergence of Trusted Execution Environments (TEEs) opens new perspectives in solving the critical and conflicting challenge of securing users’ data while enabling a rich ecosystem of data-driven applications. In this paper, we propose a PDMS architecture leveraging TEEs as a basis for security. Unlike existing solutions, our architecture allows for data processing extensiveness through the integration of any user-defined functions, albeit untrusted by the data owner. In this context, we focus on aggregate computations of large sets of database objects and provide a first study to mitigate the very large potential data leakage. We introduce the necessary security building blocks and show that an upper bound on data leakage can be guaranteed to the PDMS user. We then propose practical evaluation strategies ensuring that the potential data leakage remains minimal with a reasonable performance overhead. Finally, we validate our proposal with an Intel SGX-based PDMS implementation on real data sets.

Over the past decade, successive steps have been taken to empower individuals with new legal and technical means, from smart disclosure initiatives to the right to data portability of the GDPR and the new notion of data altruism enacted in the EU [5]. The ultimate goal is to enable individuals to collect and share their personal information, for their own good and broader societal benefits, unlocking innovative usages when crossing multiple data sources from one or many users. The technical corollary of this movement is the emergence of personal data management systems (PDMS) that allow individuals to assemble their personal data under their control, with products such as Digi.me, Cozy Cloud or Solid/PODS, as well as initiatives like Mydata.org, supported by data protection agencies.