Formal Privacy Guarantee in Predictive Autoscaling by Differentially Private Federated Training
Publication Date
3-9-2026
Document Type
Conference Proceeding
Publication Title
2026 International Conference on Computing Networking and Communications Icnc 2026
DOI
10.1109/ICNC68183.2026.11416864
First Page
302
Last Page
308
Abstract
Modern resource management in networks and cloud systems increasingly relies on machine learning. Predictive autoscaling often centralizes fine-grained logs, exposing membership and traffic-profile risks. We study trainingtime privacy and robustness for autoscaling forecasters without raw-log sharing by combining client-side differentially private stochastic gradient descent (DP-SGD) with federated aggregation. We report composed privacy at the end of training (ϵ=13.76, δ=10-5) and at the best-validation checkpoint (ϵ=9.69, δ=10-5). We audit membership inference using loss-threshold and confidence attacks, analyze heterogeneity with shard-size weighting, shard filtering, and advanced baselines (FedProx, SCAFFOLD), and quantify robustness to adversaries, including ρ=0.3 Byzantine clients (median, trimmed mean, Krum) and multiple malicious-aggregator variants. Runtime privacy is addressed in a companion systems paper [1]. On GCT v3, a DPLSTM remains operationally useful at moderate privacy budgets, reducing membership-inference advantage to about 2-3% on machines excluded from training.
Keywords
Autoscaling, Differential Privacy, Federated Learning, Heterogeneity, Membership Inference
Department
Computer Science
Recommended Citation
Alan Chuang, Melody Moh, and Teng Sheng Moh. "Formal Privacy Guarantee in Predictive Autoscaling by Differentially Private Federated Training" 2026 International Conference on Computing Networking and Communications Icnc 2026 (2026): 302-308. https://doi.org/10.1109/ICNC68183.2026.11416864
