Anomaly Detection and Debugging by Kernel Ridge Reconstruction

Abstract

AbstractWe developed an explainable reconstruction-based anomaly detection method called kernel ridge reconstruction (KRR). KRR replaces the non-linear decoder of an autoencoder (AE) with a linear regressor trained to reconstruct only normal data. Thanks to the linearity, its reconstruction can be written as the weighted sum of the normal training data, which prevents the reconstruction of anomalous data, and these weights help us to verify undesirable properties of the encoder and the training data. This linearity also reveals the desired property of the encoder for KRR to achieve high anomaly detection performance. We propose a training algorithm for the encoder to realize this property and an efficient data cleansing method. Our experiments on the MNIST, CIFAR10, and KDDCup99 datasets showed higher performance and lower computational cost of KRR than a recent reconstruction-based anomaly detection method (MemAE). The weights of the training data revealed that the AE’s encoder reconstructed anomalous data mainly composed of visually similar training data and our training algorithm prevented this problem. Our data cleansing method correctly removed harmful data from the training data and training on this cleaned dataset improved the performance of KRR.