A Lightweight Deep Autoencoder-Based Approach for Unsupervised Anomaly Detection

An Intrusion Detection System for the Internet of Things Based on the Ensemble of Unsupervised Techniques

“…Here, AE is an unsupervised neural network, which network structure is the same as UTEN-IDS. We use one statistical approach used in [42] to set the threshold of AE reconstruction loss, and the threshold is used for classification. As mentioned earlier, Kitsune is an online method and KitNET is the core detection algorithm of Kitsune.…”

Tuning a variational autoencoder for data accountability problem in the Mars Science Laboratory ground data system

“…can be applied on unsupervised anomaly detection problems [2,22,26] but they often fall short when dealing with high dimensional data with complex structures contrary to deep learning algorithms. In a semi-supervised or unsupervised context, generative neural networks such as AEs can be adapted for anomaly detection problems [21,27,28] by training them on normal data so that they learn to reproduce or generate good behaviors with a small reconstruction error. During this training, each data point is first reduced to a lower dimension representation that is expanded to its original size afterwards.…”

“…The usual framework for anomaly detection with AEs is to collect the reconstruction errors for all points of the dataset and find a threshold value that will determine which errors are considered outliers. AEs have proven themselves to be efficient tools for unsupervised anomaly detection [21,27] with the caveat that the architecture and hyperparameters of such neural networks must be adequately chosen to get a competitive performance for real life applications.…”