Skip to Content
SensorsSensors
  • Article
  • Open Access

18 November 2026

30 Pages

Beyond Latency: Deployment Assessment of Homomorphically Encrypted Inference in Sensor–Gateway–Cloud Systems

,
and
1
Department of AI, Gachon University, Seongnam-si 13120, Republic of Korea
2
Department of Electronic Engineering, Gachon University, Seongnam-si 13120, Republic of Korea
3
Department of Smart Security, Gachon University, Seongnam-si 13120, Republic of Korea
*
Authors to whom correspondence should be addressed.
This article belongs to the Section Sensor Networks

Abstract

Homomorphic encryption lets a cloud server run a model on encrypted sensor data, but fast implementations are not necessarily deployable. They may reveal intermediate values to the gateway, need the gateway to answer during inference (callbacks), or require gigabytes of keys before the first request. We propose a measurement-based deployment assessment in which admission precedes comparison: only implementations that meet service requirements for gateway exposure, callbacks, numerical fidelity, security, and resources are ranked by cost, with every exclusion recorded. We apply it to five OpenFHE implementations of LeNet and two UFPR-AMR meter-reading models over emulated broadband links. Single metrics misled: in a rejected setting, correct labels hid wrong encrypted quantization codes, and bootstrapping turned 12 requested rotations into 103 provisioned keys. Gateway-side nonlinear processing is fastest but requires disclosure and callbacks. Without either, a leveled implementation (no ciphertext renewal) qualifies under the no-intermediate-exchange premise; a bootstrapped implementation is reported only as a measured reference. On the meter-reading models, only gateway-side processing completed the assessment, at minutes per reading. All measurements used workstation-class hosts and do not establish suitability for resource-constrained gateways; energy was not measured. Deployability should be checked before latency; the assessment names the missing capability when none qualifies.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.