Effective Network Vulnerability Assessment through Model Abstraction

A significant challenge in evaluating network security stems from the scale of modern enterprise networks and the vast number of vulnerabilities regularly found in software applications. A common technique to deal with this complexity is attack graphs, where a tool automatically computes all possible ways a system can be broken into by analyzing the configuration of each host, the network, and the discovered vulnerabilities. Past work has proposed methodologies that post-process “raw” attack graphs so that the result can be abstracted and becomes easier for a human user to grasp. We notice that, while visualization is a major problem caused by the multitude of attack paths in an attack graph, a more severe problem is the distorted risk picture it renders to both human users and quantitative vulnerability assessment models. We propose that abstraction be done before attack graphs are computed, instead of after. This way we can prevent the distortion in quantitative vulnerability assessment metrics, at the same time improving visualization as well. We developed an abstract network model generator that, given reachability and configuration information of a network, provides an abstracted model with much more succinct information about the system than the raw model. The model is generated by grouping hosts based on their network reachability and vulnerability information, as well as grouping vulnerabilities with similar exploitability. We show that the attack graphs generated from this type of abstracted inputs are not only much smaller, but also provide more realistic quantitative vulnerability metrics for the whole system. We conducted experiments on both synthesized and production systems to demonstrate the effectiveness of our approach.