Scalable attack modelling in support of security information and event management

Abstract

While assessing security on single devices can be performed using vulnerability assessment tools, modelling of more intricate attacks, which incorporate multiple steps on different machines, requires more advanced techniques. Attack graphs are a promising technique, however they face a number of challenges. An attack graph is an abstract description of what attacks are possible against a specific network. Nodes in an attack graph represent the state of a network at a point in time while arcs between nodes indicate the transformation of a network from one state to another, via the exploit of a vulnerability. Using attack graphs allows system and network configuration information to be correlated and analysed to indicate imminent threats. This approach is limited by several serious issues including the state-space explosion, due to the exponential nature of the problem, and the difficulty in visualising an exhaustive graph of all potential attacks. Furthermore, the lack of availability of information regarding exploits, in a standardised format, makes it difficult to model atomic attacks in terms of exploit requirements and effects. This thesis has as its objective to address these issues and to present a proof of concept solution. It describes a proof of concept implementation of an automated attack graph based tool, to assist in evaluation of network security, assessing whether a sequence of actions could lead to an attacker gaining access to critical network resources. Key objectives are the investigation of attacks that can be modelled, discovery of attack paths, development of techniques to strengthen networks based on attack paths, and testing scalability for larger networks. The proof of concept framework, Network Vulnerability Analyser (NVA), sources vulnerability information from National Vulnerability Database (NVD), a comprehensive, publicly available vulnerability database, transforming it into atomic exploit actions. NVA combines these with a topological network model, using an automated planner to identify potential attacks on network devices. Automated planning is an area of Artificial Intelligence (AI) which focuses on the computational deliberation process of action sequences, by measuring their expected outcomes and this technique is applied to support discovery of a best possible solution to an attack graph that is created. Through the use of heuristics developed for this study, unpromising regions of an attack graph are avoided. Effectively, this prevents the state-space explosion problem associated with modelling large scale networks, only enumerating critical paths rather than an exhaustive graph. SGPlan5 was selected as the most suitable automated planner for this study and was integrated into the system, employing network and exploit models to construct critical attack paths. A critical attack path indicates the most likely attack vector to be used in compromising a targeted device. Critical attack paths are identifed by SGPlan5 by using a heuristic to search through the state-space the attack which yields the highest aggregated severity score. CVSS severity scores were selected as a means of guiding state-space exploration since they are currently the only publicly available metric which can measure the impact of an exploited vulnerability. Two analysis techniques have been implemented to further support the user in making an informed decision as to how to prevent identified attacks. Evaluation of NVA was broken down into a demonstration of its effectiveness in two case studies, and analysis of its scalability potential. Results demonstrate that NVA can successfully enumerate the expected critical attack paths and also this information to establish a solution to identified attacks. Additionally, performance and scalability testing illustrate NVA's success in application to realistically sized larger networks.