A Lightweight Deep Autoencoder-Based Approach for Unsupervised Anomaly Detection
Gcinizwe Dlamini, Rufina Galieva, Muhammad Fahim · 2019
Unsupervised anomaly detection is an important area of research to find abnormal behavior and integral part of many systems. In this research, a lightweight deep autoencoder based approach is presented to detect anomalies in unsupervised manner. It has the ability to learn the model over the normal patterns and any deviation is considered as an anomaly. Consequently, it can relax the condition to have anomalous data patterns during the training phase of the model. In this work, we examine lightweight autoencoder for anomaly detection task in order to show that simple architecture can show good performance in terms of training, testing time, number of parameters and metrics. We apply autoencoder for binary classification problem (i.e., each data point considered either normal or abnormal). The reconstruction error is used to detect anomalies. The experiments are carried out over the particular class of cyber security domain known as intrusion detection systems. We evaluated our model on standard publicly available benchmarks of KDD-99, NSL-KDD and UNSW-NB15 and achieved F1-score of 0.96, 0.88 and 0.95, respectively. It outperforms by a considerable margin when compared to state-of-the-art methods.