Empowering Network Security with Autoencoders

In the last decade, the dramatic expansion of interconnected devices, such as those in IoT and Cloud networks, has underscored the critical need for robust network security assessments. A key strategy in addressing this challenge is the detection of network anomalies, which are vital indicators of network intrusions, privacy breaches, system damage, and fraudulent activities. Deep neural networks, particularly effective in learning complex anomaly patterns in a data-driven manner, have gained significant attention in this realm. However, the effectiveness of these techniques is often limited by the unique properties of network traffic data, characterized by its sparsity and noise — attributable to the vast number of devices and diverse internet applications generating this data — and a high degree of imbalance, with anomalies typically occurring in only 0.001-1% of instances. In our work, we tackle these challenges with a streamlined approach. First, we employ an autoencoder (AE) to pinpoint instances of anomalous behavior. Subsequently, these anomalies are categorized by an attack classifier according to their specific type. Our framework has been tested on a large-scale dataset comprising real-world network traffic data, showing promising results.