Privacy-Preserving Federated Learning via Adaptive Gradient Perturbation in Edge Computing Networks

Dr. Evelyn Author [DEMO]¹*iD,Dr. Liam Zhang [DEMO]²iD

1.University of Toronto [DEMO], Canada

2.McGill Computing Lab [DEMO], Canada

DOI: 10.5555/IJCSDI.2027.0001

Assigned DOI Placeholder (Awaiting Crossref sync)

Publication CertificateDownload PDF

Received:Oct 10, 2026

Revised:Nov 25, 2026

Accepted:Dec 18, 2026

Published:Jan 15, 2027

Abstract

Federated learning enables decentralized model training across distributed clients without exposing raw local data. However, vulnerability to gradient inversion attacks remains a primary bottleneck in edge deployments. In this paper, we propose an adaptive gradient perturbation mechanism with dynamic differential privacy guarantees. Experiments demonstrate a 41% reduction in privacy leakage with less than 1.2% degradation in model convergence across standard vision benchmarks.

Keywords

Federated LearningEdge ComputingDifferential PrivacyGradient InversionNeural Optimization

Article Text

Distributed machine learning paradigms have revolutionized how client devices collaborate on neural network optimization without centralized data pooling...

References

[1] H. Brendan McMahan et al., "Communication-Efficient Learning of Deep Networks from Decentralized Data," AISTATS, 2017.
[2] L. Zhu, Z. Han, and S. Han, "Deep Leakage from Gradients," NeurIPS, 2019.
[3] C. Dwork, "Differential Privacy: A Survey of Results," TAMC, 2008.
[4] K. Bonawitz et al., "Towards Federated Learning at Scale: System Design," SysML, 2019.

Declarations & Disclosures

Funding:Funded in part by the Advanced Computing Initiative grant #ACI-2026-99.

Conflict of Interest:The authors declare no competing financial interests.

Data Availability:Benchmark scripts and differential privacy implementations are publicly archived under MIT license.