Research Paper
Zero-trust node security for KubeEdge
Abstract
A node that keeps proving itself
Industrial digital twins (DTs) increasingly run on edge nodes that ingest sensor data, update physical-asset state, and relay control or federated-learning (FL) updates across cloud–edge boundaries. If one of these hosts is compromised, the DT can continue to appear healthy while corrupting state synchronization, injecting falsified readings, or leaking proprietary FL gradients. Containerized DT workloads are commonly deployed through cloud–edge orchestration stacks, and KubeEdge is a Cloud Native Computing Foundation (CNCF) graduated framework for extending Kubernetes to intermittently connected edge environments. Its security model, however, still inherits Kubernetes’ trust-once-admit-forever assumption: after an edge node joins the cluster, no hardware-rooted evidence is required to keep justifying its trustworthiness, and cloud-side policy changes cannot be enforced while the cloud–edge link is severed.
We derive four node-integrity requirements for DT edge deployments: continuous attestation of long-lived workloads, fail-close enforcement during partitions, workload-level isolation of FL traffic, and policy enforcement across multiple trust domains. We then implement a dual-loop zero-trust architecture for KubeEdge using Keylime virtual-TPM (vTPM) remote attestation backed by the Linux Integrity Measurement Architecture (IMA), Kyverno policy-as-code admission control, and Cilium eBPF network microsegmentation. The main new component is an edge-side enforcement daemon that polls the Keylime verifier over mutual TLS (mTLS) and installs iptables FORWARD DROP rules when attestation fails or the verifier becomes unreachable, providing a local enforcement path for FORWARD-path pod traffic during cloud–edge partitions.
We evaluate the architecture with an automated 900-run matrix (3 configurations × 6 threat scenarios × 50 repetitions per cell) and report two outcome dimensions separately. (i) No security-mechanism bypass was observed within the tested threat model, where the corresponding mechanism was enabled. The exercised mechanisms were vTPM/IMA attestation (proposed only), Kyverno admission (baseline-B and proposed), and Cilium default-deny policy (baseline-B and proposed). In the proposed configuration, no untrusted workload was admitted, no Keylime-detected integrity violation went undetected, and no prohibited lateral connection succeeded; baseline-A/B S2/S4 results reflect the absence of attestation rather than an attestation success.
(ii) The end-to-end scenario-completion rate is 91.33% (822/900). In 7 S2-Proposed runs the cloud-loop network-revocation chain (CiliumNetworkPolicy delete) did not complete within the scenario’s time budget, leaving a transient post-FAIL connectivity window. The remaining 71 non-completing runs concentrate in S5 post-reconnect scheduling convergence. Both failure classes are traced to KubeEdge platform limitations, including two previously undocumented bugs. Median IMA tamper-isolation latency is 3.19 s, and normal-path admission overhead is below 1% versus the unprotected baseline. These results show how hardware-rooted, attestation-aware policy enforcement can support DT runtimes and decentralized-intelligence pipelines on mid-to-high-end industrial edge nodes without materially increasing normal-path admission latency.
In Practice
From the paper to the gateway
The zero-trust architecture behind TrustIoT.AI’s gateways comes from this research. Every factory gateway carries an identity bound to its hardware that cannot be forged or copied, so each analysis and energy report TrustIoT.AI produces rests on data from a verified device.
Full Text
Read the paper in the journal
Zero-trust node security for KubeEdge: A policy-as-service architecture and reliability evaluation. Published online on 18 September 2026 in Future Generation Computer Systems, Elsevier, volume 187 (February 2027).