ABSTRACT
Abstract The next generation of internet and wireless networks demand pervasive
storage graphs. Storage Intrusion Prevention Systems (SIDSs) are known techniques for identifying attacks. In pervasive un-directional graphs or intensive algebraic operations, victim states in final stages tend to be easily identified within a finite-time horizon by sensing abrupt transitions in system and network state spaces. However, in the early stages of such attacks, these changes are hard to prevent and difficult to distinguish from the usual state fluctuations. Most of the traditional prevention techniques employ parametric model approaches. Recently proposed model-free approximation techniques result in intensive computation. In order to provide autonomy to various large, or high dimensional state spaces, we divide the problem of attack classification into two parts: class prediction problems and cluster prediction problems. We propose a formal model-free prevention technique. Also, we propose using an autonomous utility framework in combination with a programable architecture, a set of algorithms and optimization methods which are grounded from Reinforcement Learning (RL) and Dynamic Programming (DP) techniques to tune prevention networks. We propose an autonomous learning network adjusted by the RL algorithms for stochastic intrusion environments in order to derive optimization procedures. Such procedures optimize policies and reduce time and space complexity within large prevention spaces. Moreover, we use attribute and instance selection for data reduction of pervasive learn-
approach.
