Outsourcing Computation Of Privacy Preserving Anomaly Detection Algorithm Based On Secure Multiparty Computation

Anomaly detection refers to the algorithm to find the anomalies among the data.As a branch of data mining,it has important research significance.With the rapid development of information technology,the data source of the detection algorithm is more and more diverse.In order to ensure that the data owner's privacy data is not disclosed in the process of anomaly detection,the privacy preserving scheme is necessary.Privacy preserving data mining refers to the use of data perturbation,data reconstruction,cryptography and other technical methods to ensure the accuracy of data mining results and protecting the data owner's privacy.Data perturbation and data reconstruction technology will affect the accuracy of data mining.And secure multi-party computation(SMC)method based on cryptography can ensure the accuracy of data mining,but SMC always involves a great deal of calculation under ciphertext.The efficiency of the whole method is too low to be practical.Based on the privacy-preserving requirements,this paper proposes an anomaly detection algorithm based on secure multi-party computation.This method supports the outsourcing of computation to server-side to improve the efficiency of the algorithm.In anomaly detection area,the isolation forest(i Forest)is a state of art method.The algorithm makes use of the anomalies is "less and different",using isolation tree to isolate samples.A sample is more likely to be anomalies if it has a lower height in The isolation tree.Because it indicate that the sample is easy to be segmented.At the same time,i Forest also uses the ensemble learning method.Building a series of isolation trees randomly to make up isolated forest,and uses the average depth of the sample to represent the anomaly degree.The lower the average depth,the higher the likelihood of the sample becoming anomalies.But the i Forest will face horrible security problem if we extend it to multi-party situation,because the structure of the tree is highly correlated with the dataset.So if we send this tree to others,it will expose some information about our own data which is not allowed in SMC.Aiming at this problem,we proposed an improved scheme namely security isolation forest(SIF).This method covers the original dataset information in the processing of building trees.The construction is not a random tree but a full binary tree,each node contains a variable namely size which represent the number of samples that match the node in the original dataset,and the size is encrypted by homomorphic encryption.When this full binary tree is sent out,it does not leak any sensitive information from the original dataset because all leafs are in same depth and the size was encrypted.In order to solve the problem of efficiency of secure multi-party computation,this paper uses the outsourcing computing technology to reduce the data owners computation.There are two frameworks for outsourcing computing,one is outsourcing computing and storage,the other is just outsourcing computing.This paper takes number two framework,only outsourcing computing.D ata owners are training anomaly detection model on their own dataset,then encrypted the model and broadcast it to the other party.When other party received detection models,they will get the anomaly detection result with the help of outsourcing computing server,Due to the powerful outsourcing computing server,the data owners only need a small amount of computation.Finally,this paper uses a series of datasets to verify the SIF method,and compared with some representative anomaly detection algorithms which was proposed in recent years.The effectiveness of SIF is proved through security analysis and comparative experiments.