Explainable Artificial Intelligence for Interpretable Data Minimization

Black box models such as deep neural networks are increasingly being deployed in high-stakes fields, including justice, health, and finance. Furthermore, they require a huge amount of data, and such data often contains personal information. However, the principle of data minimization in the European Union’s General Data Protection Regulation requires collecting only the data that is essential to fulfilling a particular purpose. Implementing data minimization for black box models can be difficult because it involves identifying the minimum set of variables that are relevant to the model’s prediction, which may not be apparent without access to the model’s inner workings. In addition, users are often reluctant to share all their personal information. We propose an interactive system to reduce the amount of personal data by determining the minimal set of features required for a correct prediction using explainable artificial intelligence techniques. Our proposed method can inform the user whether the provided variables contain enough information for the model to make accurate predictions or if additional variables are necessary. This human-centered approach can enable providers to minimize the amount of personal data collected for analysis and may increase the user’s trust and acceptance of the system.