Abstract. Based on L-diversity multiple sensitive modules, a hierarchical multiple sensitive attributes algorithm is proposed according to least mean square criterion ---L-LMSU (L-Least Mean Square Uniqueness). The algorithm makes a hierarchical strategy according to the frequencies of the whole attributes firstly. Beyond the hierarchical strategy, the algorithm could decrease the hidden loss because of non-uniform distribution of attributes when releasing privacy data in groups. Analysis and experiments show that L-LMSU is with linear time complexity and could improve the availability of released privacy data and time performances effectively.

Beyond l-diversity model, an algorithm (l-BDT) based on state decision tree is proposed in this paper, which aims at protecting multi-sensitive attributes from being attacked. The algorithm considers the whole situations in equivalence partitioning for the first, prunes the decision tree according to some conditions for the second, and screens out the method with the least information loss of equivalence partitioning for the last. The analysis and experiments show that the l-BDT algorithm has the best performance in controlling the information loss. It can be ensured that the published data is the most closed towards the original data, so as to ensure that the published data is as useful as possible.