Problematic Projection to the In-Sample Subspace for a Kernelized Anomaly Detector

We examine the properties and performance of kernelized anomaly detectors, with an emphasis on the Mahalanobis-distance-based kernel RX (KRX) algorithm. Although the detector generally performs well for high-bandwidth Gaussian kernels, it exhibits problematic (in some cases, catastrophic) performance for distances that are large compared to the bandwidth. By comparing KRX to two other anomaly detectors, we can trace the problem to a projection in feature space, which arises when a pseudoinverse is used on the covariance matrix in that feature space. We show that a regularized variant of KRX overcomes this difficulty and achieves superior performance over a wide range of bandwidths.

[1]  John P. Kerekes,et al.  Development of a Web-Based Application to Evaluate Target Finding Algorithms , 2008, IGARSS 2008 - 2008 IEEE International Geoscience and Remote Sensing Symposium.

[2]  J. Theiler.  BY DEFINITION UNDEFINED : ADVENTURES IN ANOMALY ( AND ANOMALOUS CHANGE ) DETECTION , 2014 .

[3]  Daniel Cremers,et al.  Shape statistics in kernel space for variational image segmentation , 2003, Pattern Recognit..

[4]  Tiziana Veracini,et al.  A Locally Adaptive Background Density Estimator: An Evolution for RX-Based Anomaly Detectors , 2014, IEEE Geoscience and Remote Sensing Letters.

[5]  Nasser M. Nasrabadi.  Kernel subspace-based anomaly detection for hyperspectral imagery , 2009, 2009 First Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing.

[6]  Heiko Hoffmann,et al.  Kernel PCA for novelty detection , 2007, Pattern Recognit..

[7]  Don R. Hush,et al.  Radial kernels and their reproducing kernel Hilbert spaces , 2010, J. Complex..

[8]  Tiziana Veracini,et al.  Background Density Nonparametric Estimation With Data-Adaptive Bandwidths for the Detection of Anomalies in Multi-Hyperspectral Imagery , 2014, IEEE Geoscience and Remote Sensing Letters.

[9]  Xiaoli Yu,et al.  Adaptive multiple-band CFAR detection of an optical pattern with unknown spectral distribution , 1990, IEEE Trans. Acoust. Speech Signal Process..

[10]  Don R. Hush,et al.  Statistics for characterizing data on the periphery , 2010, 2010 IEEE International Geoscience and Remote Sensing Symposium.

[11]  D. W. Scott,et al.  Variable Kernel Density Estimation , 1992 .

[12]  Heesung Kwon,et al.  Kernel RX-algorithm: a nonlinear anomaly detector for hyperspectral imagery , 2005, IEEE Transactions on Geoscience and Remote Sensing.

[13]  E. M. Winter,et al.  Anomaly detection from hyperspectral imagery , 2002, IEEE Signal Process. Mag..

[14]  Robert P. W. Duin,et al.  Uniform Object Generation for Optimizing One-class Classifiers , 2002, J. Mach. Learn. Res..

[15]  P. Mahalanobis.  On the generalized distance in statistics , 1936 .

[16]  S Matteoli,et al.  A tutorial overview of anomaly detection in hyperspectral images , 2010, IEEE Aerospace and Electronic Systems Magazine.

[17]  E. Parzen.  On Estimation of a Probability Density Function and Mode , 1962 .