Our publications: 
Integral privacy. We have proposed this privacy model and shown how to build integrally private computations (e.g., decision trees, statistics, regression).

Torra, V., NavarroArribas, G. (2016) Integral Privacy, Proc. CANS 2016: 661669. paper / DOI here
 This is the paper in which we defined integral privacy. The original definition considered changes on databases and how to avoid that these changes are detected when we disclose the model.

Senavirathne, N., Torra, V. (2019) Integrally private model selection for decision trees, Computers and Security 83: 167181. paper / DOI
 This paper explains how to build decision trees that are integrally private. We show that when selecting a decision tree, we can select one that is not only accurate but also recurrent. This makes the decision tree stronger from a privacy perspective. This is so because there are a large number of datasets that can generate the selected tree. The paper shows that accuracy and recurrency (and thus privacy) can both taken into account.

Senavirathne, N., Torra, V. (2019) Integral Privacy Compliant Statistics Computation, Proc. DPM/CBT@ESORICS 2019: 2238. paper / DOI here
 This paper studies the case of some statistics. We explain how to compute different types of statistics that are compliant with the definition of integral privacy.

Senavirathne, N., Torra, V. (2018) Approximating Robust Linear Regression With An Integral Privacy Guarantee, Proc. PST 2018: 110. paper (DOI) here
 We studied linear regression from an integral privacy perspective. We work with robust regression.

Torra, V., NavarroArribas, G. (2018) Probabilistic Metric Spaces for Privacy by Design Machine Learning Algorithms: Modeling Database Changes, Proc. DPM/CBT@ESORICS 2018: 422430. paper (DOI) here
 In statistical and machine learning, and when comparing data protection procedures, it is usual to compare models with respect to accuracy. In this work we consider a different approach. Distance between a pair of models is based on how these models might have been generated. So, if two different models are generated using two databases that only differ in one record, they will be similar and if they are generated by two completely different databases, they will be dissimilar. This type .
