Preference-based anonymization of numerical datasets by multi-objective microaggregation

• Our method enables the data publisher to effectively control how DR and IL are aggregated for a fixed value of κ.
• The privacy and utility of conventional microaggregation methods can be significantly improved, at the same time.
• Our partition encoding is sound and complete, and efficient for large datasets.
• Our method is general and can adapt itself to different requirements and assessment measures.

Microaggregation is a statistical disclosure control mechanism to realize k-anonymity as a basic privacy model. The method first partitions the dataset into groups of at least k records and then aggregates the group members. Generally, larger values of k provide lower Disclosure Risk (DR) at the expense of increasing Information Loss (IL). Therefore, the data publisher has to set appropriate microaggregation parameters to produce a protected and useful anonymized data. Unfortunately, in the most of the conventional microaggregation methods, the only available parameter of the algorithm, i.e., k does not enable the data publisher to effectively control the trade-off problem between DR and IL. This paper proposes a novel microaggregation method to optimize information loss and disclosure risk, simultaneously. The trade-off problem is expressed and solved within a multi-objective optimization framework. The data publisher can choose a more preferred protected dataset from a set of non-dominated candidate solutions, or even direct the method toward a desired point. Experimental results show that for a fixed value of k, the proposed method can usually produce more protected and useful datasets in comparison with the conventional methods.