ADAPTIVE FEDERATED LEARNING WITH DYNAMIC QUANTIZATION AND CLIENT ANONYMIZATION FOR PRIVACY-PRESERVING IOT-ENABLED EDGE NETWORKS

Asrorov Oybek Asror O'G'li ()
Additional contact information
Asrorov Oybek Asror O'G'li: University of Economy and Pedagogy

Synoptic: International Journal of Multidisciplinary Research, vol. 1, issue 1, 69-81

Abstract: The proliferation of Internet of Things (IoT) devices in smart infrastructure has generated massive streams of decentralized data, accelerating the adoption of Federated Learning (FL) as a privacy-preserving distributed learning paradigm. However, conventional FL frameworks face significant bottlenecks in real-world edge networks: high communication overhead caused by transmitting millions of high-dimensional model weights, extreme system heterogeneity across resource-constrained edge nodes, and vulnerability to advanced privacy attacks such as gradient inversion and membership inference. To resolve these interconnected challenges, this paper proposes an Adaptive Federated Learning framework with Dynamic Quantization and Client Anonymization (AFL-DQCA). The proposed framework introduces an adaptive gradient quantization mechanism that dynamically adjusts bit-widths (1-bit to 8-bit) based on real-time channel capacity and local computing constraints, minimizing communication payload without destabilizing global model convergence. To safeguard against privacy leakage from raw gradient inspection, we integrate a lightweight cryptographic client anonymization layer using blind signatures and a decentralized proxy relay network, severing the linkable identity between model updates and specific edge nodes. We implement and evaluate AFL-DQCA on a multi-access edge computing (MEC) testbed using standard benchmarks (MNIST, CIFAR-10, and the UNSW-NB15 network intrusion dataset). Experimental results demonstrate that AFL-DQCA reduces communication bandwidth consumption by up to 78.4% compared to standard Federated Averaging (FedAvg) and outperforms state-of-the-art quantized FL variants. Furthermore, the framework maintains high classification accuracy (within 1.2% of centralized training) while providing robust defense against gradient inversion attacks up to a peak signal-to-noise ratio (PSNR) reduction threshold of 42.1 text{ dB}, confirming its suitability for secure, scalable IoT-enabled edge networks.

Keywords: Federated Learning; Edge Computing; Internet of Things (IoT); Dynamic Quantization; Client Anonymization; Gradient Inversion Defense; Privacy-Preserving AI. (search for similar items in EconPapers)
Date: 2026-04-01
References: Add references at CitEc
Citations:

Downloads: (external link)
https://lmhqneinkzpxfhahvcvx.supabase.co/storage/v ... 9534198357-69-81.pdf (application/pdf)

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:snp:journl:art-1779534202288

Access Statistics for this article

More articles in Synoptic: International Journal of Multidisciplinary Research from Synoptic Publisher
Bibliographic data for series maintained by Editor ().