Differentially Private Distance Learning in Categorical Data

Abstract Most privacy-preserving machine learning methods are designed around continuous or numeric data, but categorical attributes common in many application scenarios, including clinical and health records, census survey data. Distance-based methods, particular, have limited applicability to since they do not capture the complexity of relationships among different values a attribute. Although distance algorithms exist for may disclose private information about individual records if applied secret dataset. To address this problem, we introduce differentially family distances between any pair attribute according way co-distributed with other forming so-called context. We define variants our algorithm show empirically that approach consumes little privacy budget while providing accurate distances, making it suitable distance-based applications, such as clustering classification.