Open Access

ARTICLE

Privacy-Preserving Edge Intelligence for Speaker Verification via Uncertainty-Aware Adaptive Feature Offloading

Yongfeng Zhang^1,*, Jie Chen^2,*

1 The School of Software Engineering, East China Normal University, Shanghai, China
2 The College of Electronic Engineering, National University of Defense Technology, Hefei, China

* Corresponding Authors: Yongfeng Zhang. Email: ; Jie Chen. Email:

(This article belongs to the Special Issue: Advanced Privacy Computing for Intelligent Distributed Networks and Systems)

Computers, Materials & Continua 2026, 89(2), 59 https://doi.org/10.32604/cmc.2026.088561

Received 06 July 2026; Accepted 18 August 2026; Issue published 15 September 2026

Abstract

Speaker verification on phones, wearables, and voice-enabled Internet-of-Things gateways must balance local privacy with reliable decisions under adverse audio. Existing privacy-preserving verification schemes generally protect a fixed representation, whereas edge-offloading policies usually adapt computation without attaching an explicit feature-disclosure budget; neither line alone coordinates trial uncertainty, communication state, and privacy expenditure. This paper presents privacy-preserving uncertainty-aware adaptive feature offloading (P-UAFO), an edge-intelligence framework that keeps raw audio local and transmits only clipped, projected, quantized, and Gaussian-perturbed intermediate features when their expected benefit justifies resource cost. Its online pipeline first estimates decision uncertainty and resource state, filters infeasible actions, selects local, low-tier, or high-tier processing, and fuses the returned score only when cloud evidence is sufficiently reliable. We establish a feature-level differential-privacy guarantee, an uncertainty threshold for offloading, a variance-optimal fusion rule, and a bounded-regret result under prediction error. The evaluation uses reproducible controlled events covering clean/noisy conditions, 0.4–12 Mb/s uplinks, 20–90 ms round-trip delay, and perturbed resource predictions. The results validate the controller logic and its practical ability to avoid unnecessary feature transmission; they do not constitute a public-corpus or mobile-device benchmark, which remains necessary before real-world deployment claims can be made.

---

Keywords

---