Logo image
Back
Privacy-Preserving and Verifiable Federated Learning Framework for Biometric Authentication at the Edge
Journal article

Privacy-Preserving and Verifiable Federated Learning Framework for Biometric Authentication at the Edge

Hao Zhou, Lin Li, Hua Dai, Geng Yang, Fusen Guo and Chao Chen
IEEE Transactions on Services Computing, Vol.First online, pp.1-18
21/04/2026
DOI: https://doi.org/10.1109/TSC.2026.3686112
Appears in  Recent Faculty of Science and Engineering Publications

Metrics

1 Record Views

Abstract

federated learning edge computing biometric verification privacy-preserving
Federated learning (FL) has emerged as a pivotal privacy-preserving paradigm for training models across decentralized biometric datasets. However, integrating secure biometric verification and data integrity in FL remains challenging, especially under non-IID distributions and resource constraints in edge environments. This paper introduces BPVFL, a novel Biometric Privacy-preserving and Verifiable Federated Learning framework, which synergizes differential privacy, cryptographic signatures, and photoplethysmogram (PPG)-based biometric recognition. Distinct from prior works, BPVFL supports both verifiable model updates and privacy-preserving identity authentication through a dual-layer protection scheme. We propose a customized privacy noise mechanism tailored for biometric features and a lightweight signature protocol to trace and validate local contributions. Experimental evaluations on three biometric datasets (SigD, BIDMC, TBME) demonstrate that BPVFL achieves over 92.7% accuracy under strong privacy guarantees, while reducing communication costs via deep network modularization. To substantiate verifiability beyond provenance, we further integrate an optional Correctness Verification Layer (CVL) based on remote attestation, which makes accepted updates protocol-correctness verifiable (i.e., generated by an approved clipping/noising/training pipeline) at a small constant overhead. We provide an empirical comparison to representative verifiable FL designs along the axes of client cryptographic time, extra uplink, and end-to-end round latency, clarifying the engineering trade-offs for edge biometric services.

Details

Logo image