Umeå University's logo

The importance of explainable machine learning models is increasing because users want to understand the reasons behind decisions in data-driven models. Interpretability and explainability emerge from this need to design comprehensible systems. This paper focuses on privacy-preserving explainable machine learning. We study two data masking techniques: maximum distance to average vector (MDAV) and additive noise. The former is for achieving k-anonymity, and the second uses Laplacian noise to avoid record leakage and provide a level of differential privacy. We are interested in the process of developing data-driven models that, at the same time, make explainable decisions and are privacy-preserving. That is, we want to avoid the decision-making process leading to disclosure. To that end, we propose building models from anonymized data. More particularly, data that are k-anonymous or that have been anonymized add an appropriate level of noise to satisfy some differential privacy requirements. In this paper, we study how explainability has been affected by these data protection procedures. We use TreeSHAP as our technique for explainability. The experiments show that we can keep up to a certain degree both accuracy and explainability. So, our results show that some trade-off between privacy and explainability is possible for data protection using k-anonymity and noise addition.

There is an increasing need to provide explainability for machine learning models. There are different alternatives to provide explainability, for example, local and global methods. One of the approaches is based on Shapley values. Privacy is another critical requirement when dealing with sensitive data. Data-driven machine learning models may lead to disclosure. Data privacy provides several methods for ensuring privacy. In this paper, we study how methods for explainability based on Shapley values are affected by privacy methods. We show that some degree of protection still permits to maintain the information of Shapley values for the four machine learning models studied. Experiments seem to indicate that among the four models, Shapley values of linear models are the most affected ones.