Cloud computing systems subject to co-resident attacks are modeled.
Balance between data security (theft) and reliability (corruption) is addressed.
Optimal data partition policy problem is formulated and solved.
Influence of cloud system parameters on the partition policy is demonstrated.
This paper models cloud computing systems subject to co-resident attacks, where an attacker can get access to a user's sensitive data through co-residence of their virtual machines on the same physical server. Both attackers’ and users’ virtual machines are distributed among cloud servers at random. It is assumed that attacker's successes in getting unauthorized access to data in different servers are independent events that can occur with a given probability. To mitigate effects of the co-resident attacks, a data protection policy based on the partition technique is applied where sensitive data are divided and distributed among multiple virtual machines in the cloud. As the information is useful only in its integrity, the attacker should get access to all of the separated data blocks to steal the information. On the other hand, corrupting any block can destroy the information and make it useless. Hence, creating more blocks can make data more difficult to steal (lower data theft probability), but easier to corrupt (higher data corruption probability). This work makes original contributions by formulating and solving constrained optimization problems to balance the data theft and data corruption probabilities. Particularly probabilistic models are first presented, which derive probabilities that an attacker can succeed in the data theft and data corruption. Further an optimal number of different data blocks (corresponding to the number of user's virtual machines) is obtained, which minimizes the data theft probability subject to meeting a data corruption probability constraint. Both fixed and uncertain numbers of attacker's virtual machines are considered. Numerical examples are presented to demonstrate influence of cloud system parameters on the optimal user's data partition policy obtained.