Privacy-Preserving Bayesian Network Learning on Distributed Heterogeneous Data
Wang Chengshan · 2007
Privacy regulations may prevent parties from sharing their data.A privacy-preserving distributed serial search algorithm of minimal description length(PP-SMDL)learning on vertically partitioned database with complete data was proposed to make parties share their data under privacy.In this method,each party owning confidential database extracted the vector for the distributed computation.Then based on Bresson,the homomorphic public key cryptosystem,private generalized dot product share protocol was deployed to compute empirical entropy of minimal description length(MDL)function,which could be further used to build the Bayesian network via distributed serial search algorithm of MDL(SMDL.) Our solution only brought forth the number of stochastic variables and their values while the results was exactly what it would be when the data were centralized.Moreover,the database could be partitioned vertically at will and the method also worked well for non-binary discrete data.The results showed that when the number of records was larger than 20 000,the normalized losses of PP-SMDL and SMDL were consistent,indicating the efficiency of the proposed method.