Evaluation of Deep Learning Approaches for Anomaly Detection

Abstract

Deep learning is a machine learning technique which is inspired by basic human instincts and functionality of the brain. It can be leveraged to tackle anomaly detection problems due to their ability in performing complex learning and prediction. However, this has been challenging due to the diversity of anomalies, class imbalance and curse of dimensionality. This research study focused on analyzing the performance of deep learning models for anomaly detection in various domains. Multi-Layer Perceptron, Deep Neural Network, Recurrent Neural Network and Auto Encoder algorithms were tested on 7 numerical datasets ranging from small scale to large scale in terms of both data size and features. The experimental design used one class classification to train the models from non-anomalous data to identify new instances as either anomalous or non-anomalous. The experimental results indicate that deep learning algorithms improve performance with the increase of data size. This study also identified certain limitations of deep learning models on anomaly detection.