A Low Inertia Guided Auto-Encoder for Anomaly Detection in Networks

In the case of One-class classification, the low variance directions tend to be more informative to build a model on target class. This paper introduces a low Inertia Autoencoder for anomaly detection. The proposed model emphasizes the low Inertia when training the network with gradient descent algorithm. The low Inertia Auto-encoder model is able to improve upon the classical Auto-encoder by incorporating a regularization term into the loss function, which is the Inertia of low dimensional data embedded in the Auto-encoder hidden layer. Experimental results, on KDDCup99 network traffic connections and Aegean WiFi Intrusion Dataset, have demonstrated clearly that our method has provided significantly better detection performance than the classical Auto-encoder. In fact, low inertia Auto-encoder allows better separation between reconstruction error distributions of normal data and anomalies.