Reconstruction based anomaly detection using distance metric

Abstract

For machine learning systems, anomaly detection which detects samples that are vastly different from the dataset is essential for achieving robust training and reliable results. Although it has been actively studied for the last several years, performing accurate detection on high-dimensional data remains a challenge. Conventional reconstruction-based methods have shown good performance, but they rely on the heuristic that reconstruction error of the anomalous data is larger than that of the normal data. Thus, these methods are not trained by optimizing a detection based objective function and show suboptimal performances. To tackle this problem, we propose an autoencoder-based method which is capable of regularizing the latent space for better detection by scoring anomalies through a distance metric. The proposed method leverages the additional distance loss function which makes the data close to the center along with the original reconstruction loss function. The effectiveness of our method is evaluated on several high-dimensional network intrusion detection datasets. Our method outperforms existing reconstruction-based methods and is also robust to hyperparameter selection.