FedCWD: Confidence-Weighted Federated Knowledge Distillation for Heterogeneous IoT Edge Devices

Abstract

Federated learning (FL) over heterogeneous IoT devices is hindered by two practical constraints: standard weightaveraging methods such as FedAvg require homogeneous model architectures across clients, and distillation-based alternatives such as FedDF depend on a labelled proxy dataset at the server. In this paper, we propose FedCWD, a federated knowledge distillation framework that eliminates both constraints. Each client trains a local model of arbitrary size, runs inference on a shared unlabelled distillation set, and uploads perclass confidence-weighted soft predictions. The server performs confidence-weighted aggregation of these predictions and distills a global student model via KL-divergence, requiring no groundtruth labels at the server. We evaluate FedCWD on the UCI-HAR human activity recognition dataset under three Dirichlet non-IID levels $(\alpha \in\{0.1,0.5,1.0\})$ and compare against four baselines: Local-Only, FedAvg, FedProx, and FedDF. FedCWD achieves 94.4% accuracy under near-IID conditions, outperforms FedDF under severe non-IID (84.3% vs. 79.0%), and reduces perround upload cost by 85.4% compared to weight-based methods, while natively supporting heterogeneous model architectures.