IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.11, November 2009 268 Manuscript received November 5, 2009 Manuscript revised November 20, 2009 Security Based Multiple Bayesian Models Combination Approach S.RAVI KISHAN 1 , S.RAJESH 2 , B N SWAMY 3 , J.V.D PRASAD 4 ABSTRACT Decision making in medical domain often involves incorporating new evidences into existing or working models reflecting the decision problems at hand. We propose a new framework that facilitates effective aggregation of multiple Bayesian Network models. The proposed framework aims to minimize time and effort required to customize and extend the original models through preserving the conditional independence relationships inherent in two or more types of Bayesian network models. We present an algorithm to systematically combine the qualitative and the quantitative parts of the different Bayesian models. Combination of Bayesian models involves integrating both structural and parameters of different models. We also describe how effective the presented algorithm and it can reduce total computational complexity Keywords: Data privacy, Bayesian networks (BN), privacy-preserving data mining. 1. INTRODUCTION A Bayesian network is a directed acyclic graph (DAG), which encodes the causal relationships between particular variables, represented in the DAG as nodes. Nodes are connected by causal links - represented by arrows - which point from parent nodes (causes) to child nodes (effects). Belief networks have been found to be useful in many applications related to reasoning and decision-making. Bayesian network (BN) is a powerful knowledge representation tool for uncertainty management and decision making. In a rapidly changing world, integrating new evidences or new fragments of knowledge in the form of multiple new models is challenging. Biomedical problems usually involve a large number of variables, complex relationships among the variables, and numerous parameters. The different evidences or models to be integrated may be from different sources, in different modeling languages, or differ in structure or in parameter, even if they may be derived from the same data sets or from experts in the same domain. Assume that a novice surgeon is planning to perform a head operation. However, he is not confident of his knowledge on nerve damnification and skin damnification. In order to make a sound decision, he needs to acquire additional knowledge related to possible nerve damnification and skin damnification in a head operation. Therefore, he seeks help from dermatology textbook and a neurology data set. Three Bayesian networks are modeled from the different sources: the dermatology textbook, the neurology data set, and the surgeon’s own domain expertise respectively. There are some common variables in all the three networks, or between only two of the networks. Combining multiple Bayesian probabilistic graphical models in a uniform manner is a tedious task; heterogamous models representing similar or overlapping pieces of information from possibly different viewpoints need to be combined both qualitatively and quantitatively. Some other efforts address topology combination in BNs, in which only two models can be combined at one time. Besides the difficulty in scaling, the resulting model can also be influenced by the order of combination, if there are more than two models to be combined. In this paper, presenting a security based Multiple Bayesian Model Combination (MBMC) framework to address both qualitative and quantitative combinations of an arbitrary number of graphical models simultaneously. 2. PROBLEM FORMULATION Consider two parties owning private data. Those parties wish to learn the Bayesian network on combination of their databases. To achieve this one party send this data to the other in encrypted form other party receives and decrypts it. The received data is merged with the local data. The resultant data is input for the BN learning process. Learning process involves 2-steps namely structure learning and parameter learning. For computation of parameters we make use of scalar product protocol to compute parameters in a secured and privacy preserved manner. Sending data from one party to another gives not only a chance to learn more information by the other party, it causes a breach for some security settings. Communication overhead is also caused because of sending full data. To overcome the above mentioned limitations, we send locally learned BN model information to other in an encrypted form. Other party receives and decrypts it.