Exploiting Kubernetes Autoscaling for Economic Denial of Sustainability

The complexity required to support the flexibility and scale of networks achievable by Kubernetes (K8s)-based applications leaves them vulnerable to Economic Denial of Sustainability (EDoS) attacks attempting to deprive targets of the financial means to sustain themselves. We develop Markov Decision Process (MDP) based models to reason about autoscaling and the components influencing scaling thresholds, and build atop the MDP a Stackelberg game structure to reason about EDoS attacks on K8s autoscaling and the circumstances under which adversaries inject traffic. Through numerical evaluations, we show examples where hourly-based charges eliminate incentives for resource scale-down, increasing attack incentive. Through the use of experiments on a realistic K8s cluster, we further show that increasing the attack intensity may render it less effective in generating surplus K8s resources compared to the increased traffic generation cost.