Research On Privacy-Preserving Bayesian Network Learning

With the spread of internet, more and more activities on data sharing and exchanging are coming forth in the network. Government departments, commercial organizations and other parties may wish to benefit from cooperative use of their data. To the contrary, privacy regulations and other privacy concerns may prevent the parties from sharing their data. Privacy-preserving distributed data mining emerges as the times require. In this paper, privacy-preserving Bayesian Network (BN) learning is studied technically,and some result are obtained as follows:.Based on of known BN learning methods, with considering the concepts, protocols and algorithms of Secure Multiparty computation (SMC), an idea of learning BN structure and parameters on the vertical or horizontal database in a privacy-preserving way is proposed, which help to achieving privacy-preserving distributed data mining.For privacy-preserving BN learning on the horizontally partitioned database with complete data, the paper puts forwards and realizes PPHC-TPDA (privacy-preserving TPDA learning on horizontally partitioned database with complete data) method using information theoretic analysis. In this method, according to the secure statistic protocol of oriented edges, we can obtain the statistical number of structure edges from every site. Then the arithmetic mean of local mutual information, regarded as the global mutual information, is applied in the BN learning. Experimental results show the good performance and high efficiency of this method.For privacy-preserving BN learning on the vertically partitioned database with complete data, the paper proposes PPVC-SMDL (privacy-preserving Serial search algorithm of MDL learning on vertically partitioned database with complete data) method basing on a serial search algorithm of MDL. In this method, private generalized scalar product share protocol is deployed to compute generalized scalar product and empirical entropy, which can be used in the process of constructing network. We can obtain the network parameters while learning structure. Compared with known methods, it shows better performance,...