Investigating the Effectiveness of Recurrent Neural Networks for Network Anomaly Detection

Network security has assumed utmost significance in today's highly linked world, where enterprises face more complex cyber threats. A key role in detecting and mitigating these threats is played by network anomaly detection, which flags odd patterns and behaviors in network data. Traditional anomaly detection techniques, overburdened with having to cope with constantly shifting threats, have inspired a search for machine learning tools like Recurrent Neural Networks (RNNs). This paper aims at exploring how well Recurrent Neural Networks perform in the field of computer network-anomaly detection. We plan a complete analysis to compare the efficacy of RNN-based models against traditional methods like statistical techniques and simple neural networks, based on an extensive set of network traffic. In the collection are both normal network traffic and malicious intrusions. Our experiments indicate that RNNs have great potential for finding anomalies in networks. Models such as these can make use of the sequential dependencies existing in network traffic data, for instance--observing anomalies that would otherwise have been overlooked. We explore various RNN architectures, hyperparameter settings, and feature representations to further improve results. This paper examines problems such as model interpretability, scalability and computing resources which severely limit the practical usability of RNNs in real-world network security situations. We also provide means of making RNN-based anomaly detection systems immune to malicious interference.