Meeting SLO with Privacy: Partial Homomorphic Encryption Inference and Anomaly Gating for Predictive Autoscaling in Cloud Environments
Publication Date
3-9-2026
Document Type
Conference Proceeding
Publication Title
2026 International Conference on Computing Networking and Communications Icnc 2026
DOI
10.1109/ICNC68183.2026.11416887
First Page
756
Last Page
762
Abstract
Cloud autoscaling dynamically provisions resources under fluctuating demand, but predictive autoscaling must also meet strict service-level objectives (SLO) while protecting tenant data confidentiality. In this paper we present SecurePredictScale, a runtime-secure autoscaling framework that combines DPLSTM forecasting, a CKKS-encrypted fully connected layer, inline anomaly detection (Isolation Forest) to filter poisoned and OOD inputs, and OPA-enforced RBAC. The design exposes a minimal homomorphic-encryption boundary and quantifies SLO compliance under privacy constraints. On Google Cluster Trace v3 and a containerized testbed, the encrypted FC layer completes in ≈35ms, and the end-to-end API → action loop stays under 50 ms (p95) at up to 82 requests/s. The anomaly gate reaches ≈97% TPR at ≈1% FPR, and RBAC blocks 100% of unauthorized requests. DP-SGD training achieves composed privacy ϵ ≈ 9.7 at δ=10-5 while keeping forecasting error within operational bounds and outperforming ARIMA and Prophet at comparable privacy budgets, yielding a practical template for privacy-aware autoscaling and other ML-driven control loops in cloud and edge environments.
Keywords
Anomaly Detection, Cloud Autoscaling, Differential Privacy, Edge Computing, Homomorphic Encryption, RBAC
Department
Computer Science
Recommended Citation
Alan Chuang, Melody Moh, and Teng Sheng Moh. "Meeting SLO with Privacy: Partial Homomorphic Encryption Inference and Anomaly Gating for Predictive Autoscaling in Cloud Environments" 2026 International Conference on Computing Networking and Communications Icnc 2026 (2026): 756-762. https://doi.org/10.1109/ICNC68183.2026.11416887
