Multi-paradigm frameworks for scalable intrusion detection

Research in network security and intrusion detection systems (IDSs) has typically focused on small or artificial data sets. Tools are developed that work well on these data sets but have trouble meeting the demands of real-world, large-scale network environments. In addressing this problem, improvements must be made to the foundations of intrusion detection systems, including data management, IDS accuracy and alert volume. We address data management of network security and intrusion detection information by presenting a database mediator system that provides single query access via a domain specific query language. Results are returned in the form of XML using web services, allowing analysts to access information from remote networks in a uniform manner. The system also provides scalable data capture of log data for multi-terabyte datasets. Next, we address IDS alert accuracy by building an agent-based framework that utilizes web services to make the system easy to deploy and capable of spanning network boundaries. Agents in the framework process IDS alerts managed by a central alert broker. The broker can define processing hierarchies by assigning dependencies on agents to achieve scalability. The framework can also be used for the task of event correlation, or gathering information relevant to an IDS alert. Lastly, we address alert volume by presenting an approach to alert correlation that is IDS independent. Using correlated events gathered in our agent framework, we build a feature vector for each IDS alert representing the network traffic profile of the internal host at the time of the alert. This feature vector is used as a statistical fingerprint in a clustering algorithm that groups related alerts. We analyze our results with a combination of domain expert evaluation and feature selection.