Application Specific Database Intrusion Detection Using Data Mining Techniques

Abstract

There is a growing concern about the risk exposure of databases due to massive usage of web-based applications. However, this has proliferated various forms of application level attacks causing huge financial losses to organizations. Therefore, accurate detection of fraud is inevitable in application-specific domains. In this thesis, we initially present a mobile phone fraud detection system using Fuzzy C- Means clustering and Support Vector Machine. The proposed model works by initially building calling profiles of subscribers. For identifying the illegitimate calls, trained Support Vector Machine classifier is applied. Furthermore, an alternative approach using an unsupervised Quarter-Sphere Support Vector Machine has been suggested to identify the fraudulent calls. It is observed that the Fuzzy C-Means clustering does not generate clusters quite well due to random initialization of cluster centers. Therefore, the proposed fraud detection model is improvised by applying Genetic Algorithm on Fuzzy C-Means for getting optimized clusters. However, it is found that the performance of optimized fuzzy clusters are affected by the presence of noisy points, while the efficiency of Support Vector Machine is limited due to the huge computations required for its kernel function. Hence, another telecom fraud detection system has been proposed that uses Possibilistic Fuzzy C-Means for clustering and Hidden Markov Model for better classification. After successfully developing telecom fraud detection models, we have investigated the problem of detecting fraudulent claims in automobile insurance domain. In this domain, the class imbalance problem is prevalent and affects the performance of detection system, and thus, needs to be addressed first. Therefore, we have suggested different automobile insurance fraud detection models for reducing data skewness and thereby, improving the model performance. Initially, an undersampling method that makes use of Genetic Algorithm based Fuzzy C-Means clustering has been applied. Once the