The privacy preserving data mining is a research hotspot. Most of the privacy preserving algorithms are focused on the centralized database. The algorithms on the distributed database are very vulnerable to collusion attack. This paper proposes a modified private preserving SVM (MPPSVM) classification algorithm on distributed database. In which, the SVM is used for classification on distributed database, and the gram matrix is used for creating the global model of SVM. To preserve data privacy, data on each party is split into k slices, and each party keeps one slice as private, sends the remaining slices to the other k-1 parties. Each party sends k-1 slices to k-1 parties one for each, and receives k-1 slices from others. Each party adds the k-1 slices to its private slice, and sends the data to data center. The data center receives the data, and sums them up to get the global model. This algorithm can protect the data from disclosing to other parties of the distributed database. If the data center does not take part in the collusion attack, others can't get the exact data of a party. If the data center attends the collusion attack, the disclosure possibility is very small when the collusion number is small. With the increasing number of the collusion, the disclosure possibility may be increased, but the security can still be guaranteed by changing the coefficient. We did some experiments on the algorithm, and the results show that the algorithm protects the data privacy better than other algorithms.
展开▼
