Dependence-aware modeling of multi-tenant attacks in cloud systems

Abstract

Multi-tenant cloud systems are increasingly vulnerable to co-residency attacks, in which adversaries deploy attacker virtual machines (AVMs) to compromise service component versions (SCVs) colocated on shared physical servers. Conventional reliability models often assume independent SCV failures, overlooking dependencies arising from shared vulnerabilities or coordinated attacks. We introduce a dependence-aware probabilistic framework that explicitly models statistical dependence among SCV compromises via copula-based joint distributions, and incorporates various AVM placement policies (random, hash, affinity). We analyze how SCV dependence structure, the number of attacker accounts, and IDS detection sensitivity affect the overall corruption probability. The risk model is further embedded in a Stackelberg game between defender and attacker, incorporating budget and risk-cap constraints and various detection cost regimes. We prove equilibrium existence and compute optimal strategies via a Monte Carlo procedure. It is discovered that dependence significantly increases the risk of corruption. The probability of corruption can increase by up to 75% compared to the independence baseline, with non-overlapping confidence intervals across different copula families and placement policies. Equilibrium analysis shows that placement and cost structure jointly determine the optimal detection sensitivity. These results demonstrate how dependence modeling, placement realism, and operational constraints together shape cloud service resilience and defender strategy.