Um modelo de diagnóstico distribuído e hierárquico para tolerância a ataques de manipulação de resultados em grades computacionais

Abstract

Information security in grid computing involves requirements that go beyond the traditional networks. Concerning specifically Integrity, most of existing solutions deal with it only during data transmission, ensuring non-violation of data in the communication between machines. Nevertheless, it is also necessary to guarantee data integrity during its processing, so that jobs results must not suffering any improper handling. Otherwise, the results manipulation compromises the application as a whole, also causing a high processing overhead. In order to avoid that grid users obtain uncorrected results due to malicious elements, this work proposes a diagnosis model for tolerating security faults in largescale computational grids, considering the integrity verification during job processing. In this way, it is possible to exclude those misbehaving processing units (nodes) interested in damaging the execution of processes, providing thus a high performance computing environment only formed by reliable nodes. The usage of system-level diagnosis as a strategy against jobs results manipulation attacks reveals itself an efficient solution, since it does not depend on the hardware platform and it is interoperable with security local solutions. This feature allows the employment of the proposed model at the majority of grid computing middlewares. Furthermore, the proposed diagnosis model organizes the nodes into logical clusters, establishing a hierarchy among them, in accordance to the role of each node (executor, tester or ultra-reliable), assigned through its historical behavior in the environment. This approach enables diagnosis to be made in a distributed way with the participation of nodes that provide an expected degree of confidence. To validate this strategy, a new security layer was implemented in a grid simulator in order to introduce the assertions and the behaviors described in the proposed model. The results testify the effectiveness of the model at scenarios with different quotas of malicious nodes, providing a rate of detection of 100% and accuracy of 99,7% of processed jobs, with 12,3% of overhead.