Achieving high survivability in distributed systems through automated response
We propose a new model for automated response in distributed systems. We formalize the process of providing automated responses and the criterion for asserting global optimality of the selection of responses. We show that reaching the globally optimal solution is an NP-hard problem. Therefore we design a genetic algorithm framework for searching for good selections of responses in the runtime. Our system constantly adapts itself to the changing environment based on short-term history and also tracks the patterns of attacks in a long-term history. ^ Unknown security attacks, or zero-day attacks, exploit unknown or undisclosed vulnerabilities and can cause devastating damage. The escalation pattern, commonly represented as an attack graph, is not known a priori for a zero-day attack. Hence, a typical response system provides ineffective or drastic responses. Our system “conceptualizes” nodes in an attack graph, whereby they are generalized based on the object-oriented hierarchy for components and alerts. This is done based on our insight that high level manifestations of unknown attacks may bear similarity with those of previously seen attacks. This allows the use of history such as effectiveness of each response from past attacks to assist responses to the unknown attack. ^ This thesis lays down three distinct claims and validates them empirically. The claims are: (i) For automated response, consider a baseline mechanism that has a static mapping from the local detector symptom to a local response. This corresponds to the state-of-the-art in deployed response systems. Now consider our proposed model which takes into account global optimality from choosing a set of responses and also does a dynamic computation of the response combination from the set of detectors and other system parameters (inferences about the accuracy of the attack diagnosis, response effectiveness, etc.). The survivability of the application system with our proposed model is an upper bound of the survivability achievable through the baseline model. (ii) In some practical situations, the proposed model gives higher survivability than the baseline model. (iii) The survivability with our proposed model is improved when the system takes into account history from prior similar attacks. This kind of history is particularly important when the system deals with zero-day attacks.^
Saurabh Bagchi, Purdue University.
Engineering, System Science|Computer Science