LAUSR

High-Power electric grid networks require extreme security in their associated telecommunication network to ensure protection and control throughout power transmission. Accordingly, supervisory control and data acquisition systems form a vital part of any critical infrastructure, and the safety of the associated telecommunication network from intrusion is crucial. Whereas events related to operation and maintenance are often available and carefully documented, only some tools have been proposed to discriminate the information dealing with the heterogeneous data from intrusion detection systems and to support the network engineers. In this work, we present the use of deep learning techniques, such as Autoencoders or conventional Multiple Correspondence Analysis, to analyze and prune the events on power communication networks in terms of categorical data types often used in anomaly and intrusion detection (such as addresses or anomaly description). This analysis allows us to quantify and statistically describe high-severity events. Overall, portions of alerts around 5-10% have been prioritized in the analysis as first to handle by managers. Moreover, probability clouds of alerts have been shown to configure explicit manifolds in latent spaces. These results offer a homogeneous framework for implementing anomaly detection prioritization in power communication networks.