Network Traffic Monitoring : Advanced Autoencoder Models for Real-Time Anomaly Detection

Abstract

The viability of newer versions of autoencoder models for detecting anomalies in real-time in network traffic is discussed. With the increase in systems interconnection, it becomes critically important to understand how these networks evolve and detect aberrant behavior in time. An autoencoder is a neural network frequently used in anomaly detection: The network trains itself to map a set of predetermined standard patterns. At the same time, outliers are easily recognizable since they do not fit the usual pattern. This study involves training a network traffic autoencoder to build a model of normal activity, which allows for identifying anomalies through reconstruction error. Two approaches are used to evaluate the proposed model: the batch processing approach and the streaming scenario based on real networks. The same is achieved by comparing precision, recall, and F1 scores, which explains the possibility of using autoencoders for real-time network traffic monitoring. Further presented are the problem areas concerning the threshold optimization, the system's processing power, and the possible solutions for future improvements.