Abstract—The most general mode of detecting anomalous data would make no assumptions regarding them other than their atypicality. For such a system, choosing features, to best support the detection, is problematic. The hidden layer representation of auto-encoder artificial neural networks is a potential uncommitted solution to this. We have assessed these features on a range of problems derived from two image datasets, feeding the features into one-class Radial Basis Function (RBF) -Support Vector Machine (SVM) classifiers. Our range of problems vary in diversity of the normal and anomalous classes. Assessed across the range, we find the best performing feature to be a late fusion of hidden layer activations, residual error vectors and the raw input signals. This improves upon the use of auto-encoder residual vector error magnitude, which has previously been proposed for anomaly detection.
Index Terms—Anomaly detection, auto-encoders, one-class support vector machines.
J. T. A. Andrews is with the Department of Computer Science, Department of Statistical Science, and Security Science Doctoral Training Centre, University College London, 66-72 Gower Street, WC1E 6BT, London, UK (e-mail: jerone.andrews@cs.ucl.ac.uk).
E. J. Morton is with Rapiscan Systems Ltd., 2805 Columbia St, Torrance, CA 90503 USA (e-mail: emorton@rapiscansystems.com).
L. D. Griffin is with the Department of Computer Science, University College London, 66-72 Gower Street, WC1E 6BT, London, UK (e-mail: l.griffin@cs.ucl.ac.uk).
Cite: Jerone T. A. Andrews, Edward J. Morton, and Lewis D. Griffin, "Detecting Anomalous Data Using Auto-Encoders," International Journal of Machine Learning and Computing vol.6, no. 1, pp. 21-26, 2016.