CYBER THREAT MODEL OF CONTAINERIZED CLOUD ENVIRONMENTS

Abstract

Containerised cloud environments managed by Kubernetes-class orchestrators have their own control plane: an API server, an admission service, a scheduler and scaling controllers. Each of these components is a separate attack surface, and the corresponding attacks mostly lack distinctive signs in network flows. This paper develops a formal cyber threat model of an orchestrated cloud environment that defines the infrastructure components, classifies six infrastructure-level attack vectors (container escape, privilege escalation, orchestrator API abuse, lateral movement, resource exhaustion, image supply-chain attacks) and maps them to observable telemetry features. Each vector is described by the attacker’s initial capabilities, preconditions, target components, stages and a set of features. A 51-feature space across six groups is formed, including a group of authentication features that separates privilege escalation from API abuse. Based on the model, infrastructure threat prediction is formulated as detecting attack episodes before their impact stage at a prescribed false-positive rate. The proposed HMS-AE-AT method extends SAE-AT with four new components: a multi-scale anomaly score for stealthy attacks, a buffer-poisoning-resistant adaptive threshold, a hierarchical detector that analyses the consistency of alarms across namespaces (telling planned cluster changes from attacks and detecting distributed attacks), and subject authentication features. Experiments were run on deliberately harder, non-circular cluster telemetry for 12 namespaces (5 runs; ordinary, stealthy, and distributed attacks; load drift; and a novel legitimate cluster upgrade event). The method attains the best ranking quality among semi-supervised methods (ROC-AUC 0.881 versus 0.811 for direct transfer of SAE-AT); the hierarchical detector halves the full false positive rate during a cluster upgrade (0.035 versus 0.018) with almost no loss of recall; the robust threshold prevents a slow attack from lifting the threshold (the naive threshold grows 43-fold, the robust one 1.12-fold); adding authentication features reduced the V2/V3 signature overlap from 0.43 to 0.29 and improved attribution. When a vector is absent from the training set, supervised Random Forest recall drops sharply, whereas the proposed method retains it. A trade-off is stated context-aware normalisation improves ranking and lowers the false positive rate, but weakens sensitivity to the stealthiest attacks; validation was done on emulated rather than real data.