An Unsupervised Feature Selection Method for Data-Driven Anomaly Detection Systems

摘要

Feature selection has been widely used as a pre-processing step that helps to optimise the performance of data-driven intrusion/anomaly detection systems in achieving their tasks. For example, when grouping the data into normal and outlier groups, the existence of redundant and non-representative features would reduce the accuracy of classifying the data points and would also increase the processing time. Therefore, feature selection is applied as a pre-processing step for anomaly detection systems in order to optimize their classification accuracy and running time. Most of the existing feature selection methods have limitations when dealing with high-dimensional data, as they search different subsets of features to find accurate representations of all features. Obviously, searching for different combinations of features is computationally very expensive, which makes existing work not efficient for high-dimensional data. The work carried out here, which relates to the design of a similaritybased unsupervised feature selection method for an efficient and accurate anomaly detection (UFSAD), tackles mainly the selection of reduced set of representative features from high-dimensional data without the data class labels. The selected features should improve the accuracy and performance of anomaly detection systems due to the elimination of redundant and non-representative features. The proposed UFSAD method extends the k-mean clustering algorithm to partition the features into k clusters based on a similarity measure (e.g. PCC - Pearson Correlation Coefficient, LSRE - Least Square Regression Error or MICI - Maximal Information Compression Index) in order to accurately partition the features. Then the proposed centroid-based feature selection method is used, where the feature with the closest similarity to its cluster centroid is selected as the representative feature while others are discarded. Extensive experimental work has shown that UFSAD can generate a reduced representative and non-redundant feature set that achieves good classification accuracy in comparison with well-known unsupervised features selection methods.