Learning-Based Network Intrusion Detection: an Imbalanced, Constantly Evolving and Timely Problem

Network intrusion detection systems (NIDS) observe network traffic and aim to pinpoint intrusions, i.e. effective threats on the integrity, availability or confidentiality of services and data provided by this network. There are two types of NIDS:1) signature-based intrusion detection systems that identify known intrusions by referring to an existing knowledge base, and2) anomaly-based intrusion detection systems (AIDS) that detect intrusions based on deviations from a model of normal network traffic, usually learnt through machine learning techniques.While AIDS have the advantage to not necessitate the manual creation of signatures, deploying AIDS in networks is challenging in practice.First, collecting representative network data and properly labelling it is complex and costly. This data is also highly unbalanced, as attacks are rare events. Finally, a learned AIDS is likely to show a drop in detection rates due to differences between the training context and the inference context.This presentation will discuss the results of Nicolas Sourbier’s PhD thesis that has studied how genetic programming and Tangled Program Graphs (TPGs) machine learning can help overcoming the challenges of the network AIDS.