CaBaFL: Asynchronous Federated Learning via Hierarchical Cache and Feature Balance

TL;DR

CaBaFL tackles two core FL challenges—stragglers and non-IID data imbalance—in AIoT settings by introducing a hierarchical two-level cache for asynchronous intermediate-model aggregation and a feature-balance device selection guided by middle-layer activation distributions. The method maintains multiple parallel intermediate models, dispatches them to devices, and promotes models to L1 caches for aggregation based on activation-based similarity to the global feature distribution. Empirical results on CIFAR-10/100 and FEMNIST show CaBaFL achieves substantial acceleration and accuracy gains over synchronous and asynchronous baselines, with improved stability and reduced communication overhead. The work also discusses scalability, fairness, security considerations, and behavior under different hyperparameters, and validates the approach on real hardware. Overall, CaBaFL offers a practical framework for efficient, balanced FL in large-scale AIoT deployments.

Abstract

Federated Learning (FL) as a promising distributed machine learning paradigm has been widely adopted in Artificial Intelligence of Things (AIoT) applications. However, the efficiency and inference capability of FL is seriously limited due to the presence of stragglers and data imbalance across massive AIoT devices, respectively. To address the above challenges, we present a novel asynchronous FL approach named CaBaFL, which includes a hierarchical Cache-based aggregation mechanism and a feature Balance-guided device selection strategy. CaBaFL maintains multiple intermediate models simultaneously for local training. The hierarchical cache-based aggregation mechanism enables each intermediate model to be trained on multiple devices to align the training time and mitigate the straggler issue. In specific, each intermediate model is stored in a low-level cache for local training and when it is trained by sufficient local devices, it will be stored in a high-level cache for aggregation. To address the problem of imbalanced data, the feature balance-guided device selection strategy in CaBaFL adopts the activation distribution as a metric, which enables each intermediate model to be trained across devices with totally balanced data distributions before aggregation. Experimental results show that compared with the state-of-the-art FL methods, CaBaFL achieves up to 9.26X training acceleration and 19.71\% accuracy improvements.

Figures (15)

• Figure 1: A motivating example of our asynchronous FL method.
• Figure 2: Cosine similarity comparison of activation distributions.
• Figure 3: Cosine similarity comparison on CIFAR-100 dataset.
• Figure 4: Framework and workflow of CaBaFL.
• Figure 5: Learning curves of training CIFAR-10 with ResNet-18.