Previous research proposes a random sampling-based techniques [4], [16] to produce differentially private k anonymized data based on random sampling. However, their approach considers all variables in an original dataset as quasi-identifiers with full domain transformation to guarantee differential privacy. Since all the equivalence classes in anonymized data contain identical records with the same set of attribute values, that anonymized data is not useful for the analysis of microdata. However, we consider random sampling a promising approach to guaranteeing differential privacy on frequency tables since k-anonymized data in which all variables are treated as quasi-identifiers is equivalent to a frequency table.In this paper, we evaluate the feasibility of producing deferentially private frequency tables based on random sampling. Our experiments show that we can produce frequency tables of high data utility when we set small $\epsilon$ to choose a low sampling rate for a dataset with a large number of records.

The objective of the cell suppression problem (CSP) is to protect sensitive cell values in tabular data under the presence of linear relations concerning marginal sums. Previous algorithms for solving CSPs ensure that every sensitive cell has enough uncertainty on its values based on the interval width of all possible values. However, every deterministic CSP algorithm is vulnerable to an attack scheme that narrows down the width of sensitive cell values by matching the suppression pattern of an original table with that of each candidate table with the same CSP algorithm. Although to make a CSP algorithm non-deterministic is a promising approach against the matching attack, we find that there still exists an expanded matching attack to the algorithm.

To release microdata of medical information for secondary use, it is necessary to anonymize it to protect the individual’s privacy in that data. In k-anonymity, we generalize identifying attributes of each individual to partition the data into groups of more than k records that take the same values for those identifying attributes. However, generalizing geographical information on individuals based on a domain- level hierarchy leads to anonymized data of low data utility because the population densities of regions vary significantly in Japan. Therefore, we develop a new technique of recursively partitioning regional information based on GPS coordinates. Our experimental results show that the proposed method adjusts the granularity of geographical information flexibly such that the resulting groups of records for k-anonymity possess much better uniformity in size than those with the conventional generalization method.