Generalized bucketization scheme for flexible privacy settings

Bucketization is an anonymization technique for publishing sensitive data. The idea is to group records into small buckets to obscure the record-level association between sensitive information and identifying information. Compared to the traditional generalization technique, bucketization does not require a taxonomy of attribute values, so is applicable to more data sets. A drawback of previous bucketization schemes is the uniform privacy setting and uniform bucket size, which often results in a non-achievable privacy goal or excessive information loss if sensitive values have variable sensitivity. In this work, we present a flexible bucketization scheme to address these issues. In the flexible scheme, each sensitive value can have its own privacy setting and buckets of different sizes can be formed. The challenge is to determine proper bucket sizes and group sensitive values into buckets so that the privacy setting of each sensitive value can be satisfied and overall information loss is minimized. We define the bucket setting problem to formalize this requirement. We present two efficient solutions to this problem. The first solution is optimal under the assumption that two different bucket sizes are allowed, and the second solution is heuristic without this assumption. We experimentally evaluate the effectiveness of this generalized bucketization scheme.