Collusion Detection for Grid Computing Eugen Staab and Thomas Engel Faculty of Science, Technology and Communication University of Luxembourg L-1359 Luxembourg {eugen.staab,thomas.engel}@uni.lu Abstract A common technique for result verification in grid com- puting is to delegate a computation redundantly to different workers and apply majority voting to the returned results. However, the technique is sensitive to “collusion” where a majority of malicious workers collectively returns the same incorrect result. In this paper, we propose a mechanism that identifies groups of colluding workers. The mechanism is based on the fact that colluders can succeed in a vote only when they hold the majority. This information allows us to build clusters of workers that voted similarly in the past, and so detect collusion. We find that the more strongly workers collude, the better they can be identified. 1 Introduction Problem Context and Motivation In grid computing, a master assigns computational tasks to resources identified as workers, which are expected to execute the tasks and re- turn their results. Since the master has no control over the workers, it cannot be sure whether a returned result is ac- tually correct or not. This is especially relevant in the sce- nario of desktop grids [12]. Workers may return incorrect results because they fail due to overclocking or software er- rors [13, 21], because they want to harm the master, or be- cause they want to save resources (e.g., by returning random results). Conventional security mechanisms can ensure data authenticity and integrity [20]. However, these mechanisms cannot ensure the correctness of received results. For spe- cific computations, simple checkers can be used to verify results in an efficient way (e.g., see [24]). Unfortunately, no such mechanism is known for general computations [10]. A common principle to tackle this issue in the general case is to rely on redundancy (e.g., see [1]): a computa- tion is redundantly outsourced to several randomly selected workers; majority voting is applied to the set of returned results to decide in favor of the result that appears most of- ten. This approach tolerates a certain number of incorrect results in a vote. However, it does not resist a majority of colluding workers that collectively return the same incorrect result. Even though workers are randomly selected for each vote, with the possibility of massive attacks (e.g., [4]), the probability for a majority of colluders becomes significant. Therefore, mechanisms are required that detect colluding behavior of malicious workers. Approach and Contribution We present a mechanism for collusion detection that exploits the information of how often pairs of workers are together in the majority/minority of votes, and how often they are in opposite groups. In cases, where colluding workers win a vote, they are always together in the majority, whereas honest workers together form the minority. We first show theoretically that this fact allows a line to be drawn between honest and colluding workers. Secondly, we propose an algorithm that uses graph clustering to discover this division. Finally, we evaluate the algorithm in terms of accuracy and running time, using two different graph clustering algorithms from the literature. We find that, given a certain number of observations, our mech- anism can successfully detect sophisticated colluders. Organization The remainder of this paper is organized as follows. In Sect. 2, we detail the model and assump- tions used in our work. In Sect. 3, we conduct a theoretical analysis, which forms the basis for the collusion detection algorithm proposed in Sect. 4. We describe the implemen- tation of the algorithm in Sect. 5 and evaluate its accuracy and performance in Sect. 6. We outline related work in Sect. 7 and draw conclusions in Sect. 8. 2 Model and Assumptions In this section, we first detail how the redundancy and majority voting principles are used in our work. Then, we specify the attacker models which are used in the theoretical analysis and on which our mechanism will be evaluated. 9th IEEE/ACM International Symposium on Cluster Computing and the Grid 978-0-7695-3622-4/09 $25.00 © 2009 IEEE DOI 10.1109/CCGRID.2009.12 412