AdaptiveFL: Adaptive Heterogeneous Federated Learning for Resource-Constrained AIoT Systems

TL;DR

AdaptiveFL tackles resource heterogeneity in AIoT Federated Learning by generating heterogeneous local models through fine-grained width-wise pruning controlled by $r_w$ and $I$, and by dispatching these models via an RL-based client selector that relies on historical model sizes without exposing device resources. It maintains two learning tables $T_r$ and $T_c$ to inform resource-aware and curiosity-driven decisions, and aggregates updates by aligning parameters to the full model indices with data-size weighting. The approach is validated on CIFAR-10/100 and FEMNIST, plus a real IoT test-bed, showing consistent gains over All-Large FedAvg, Decoupled FedAvg, HeteroFL, and ScaleFL, including up to around $3\%$ accuracy improvements and reduced communication waste. This work demonstrates a scalable and privacy-preserving solution for resource-constrained AIoT that adapts to dynamic hardware conditions while preserving data locality.

Abstract

Although Federated Learning (FL) is promising to enable collaborative learning among Artificial Intelligence of Things (AIoT) devices, it suffers from the problem of low classification performance due to various heterogeneity factors (e.g., computing capacity, memory size) of devices and uncertain operating environments. To address these issues, this paper introduces an effective FL approach named AdaptiveFL based on a novel fine-grained width-wise model pruning strategy, which can generate various heterogeneous local models for heterogeneous AIoT devices. By using our proposed reinforcement learning-based device selection mechanism, AdaptiveFL can adaptively dispatch suitable heterogeneous models to corresponding AIoT devices on the fly based on their available resources for local training. Experimental results show that, compared to state-of-the-art methods, AdaptiveFL can achieve up to 16.83% inference improvements for both IID and non-IID scenarios.

Figures (6)

• Figure 1: Framework and workflow of AdaptiveFL.
• Figure 2: Learning curves of AdaptiveFL and three baselines.
• Figure 3: Comparison of submodels at different levels.
• Figure 4: Learning curves on different numbers of involved devices.
• Figure 5: Ablation study for RL-based client selection.