Privacy preservation satisfying utility requirements based on multi-objective optimization

Data publishing requires privacy preservation maintaining utility. Therefore, it is necessary to balance the tradeoff between privacy (security) and utility. To achieve this, it is necessary to set parameters appropriately for data anonymization. This step usually requires trial and error by hand, but it is more convenient to automatically anonymize data for frequent data publication. We propose an automatic parameter-tuning method that maintains the required utility while preserving privacy by using multi-objective optimization with objective functions of the index to measure privacy level and utility with design parameters for anonymization. We use the non-dominated sorting genetic algorithm II to balance the above tradeoff. Experiments on diabetes data showed that our method maintains utility while preserving privacy and can obtain non-dominated solutions that balance the tradeoff. Our method includes various anonymization techniques such as top-bottom coding, k-anonymization, outlier removal, random replacement, and differential privacy.

I. Introduction

Personal data can be easily recorded with smartphones, and advances in machine learning have made big data available for service personalization. However, the collection of big data can unintentionally invade privacy, and the companies that leak privacy information are severely punished.