FedACC adapts federated learning to uneven client data with dynamic parameter splitting and clustering

A newly published framework called FedACC tackles this problem by deciding, on the fly, which parts of the model should be shared and which should stay local.

The work, published in the journal Cluster Computing, was authored by Xiaohui Yang and Jiaxiang Wu. It targets a limitation the authors identify in existing personalized federated learning (PFL) methods, which typically predefine certain layers as personalized while keeping the remaining layers shared among all participants.

According to the authors, such static partitioning struggles to adapt to evolving non-independent and identically distributed (non-IID) data. The mismatch can cut both ways: a model may become over-localized, losing the benefit of collective training, or it may suppress useful client-specific knowledge by forcing too much of the network into the shared portion.

FedACC addresses this with two cooperating modules. The first, adaptive masked personalization (AMP), dynamically separates shared and personalized parameters using two signals: gradient magnitude and gradient direction consistency. Rather than fixing the split in advance, the framework lets the data distribution itself shape which parameters remain common and which become client-specific.

The second module, adaptive hierarchical federated clustering (AHFC), groups clients according to the features of their personalized parameters and then performs hierarchical aggregation. This means clients with similar local data characteristics can be aggregated together rather than being averaged into one undifferentiated global update.

The authors evaluated FedACC on three benchmark datasets: CIFAR-10, CIFAR-10-C, and Mini-ImageNet. Across all three, the framework achieved the highest final personalized accuracy among the methods compared, improving over the strongest baseline methods by 1.23, 2.69, and 0.35 percentage points, respectively.

Beyond raw accuracy, the paper reports additional ablation studies examining the contributions of the individual modules, together with a communication-cost analysis. These experiments indicate that AMP and AHFC jointly improve personalization while reducing the number of uploaded shared parameters as the personalized subnetwork grows.

The reported gains are measured on standard image benchmarks rather than real-world deployments, and the improvements over the strongest baselines vary considerably in size, from 0.35 to 2.69 percentage points depending on the dataset. Readers should therefore treat the results as evidence of a consistent edge across these three testbeds rather than a guarantee of similar margins in every setting.

Even so, the framework’s core idea, letting the boundary between shared and personalized parameters move with the data instead of fixing it in advance, offers a concrete alternative to static layer partitioning, and the accompanying clustering step suggests how federated systems might organize heterogeneous clients without collapsing their differences into a single average model.

Subject of Research: A personalized federated learning framework combining adaptive parameter decoupling with client-side clustering to handle heterogeneous data distributions.

Article Title: FedACC: personalized federated learning via adaptive parameter decoupling and client-side clustering

Article References: Yang, X., & Wu, J. (2026). FedACC: personalized federated learning via adaptive parameter decoupling and client-side clustering. Cluster Computing, 29(16), Article 893. https://doi.org/10.1007/s10586-026-06702-9

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06702-9

Keywords: federated learning, personalized federated learning, non-IID data, parameter decoupling, client clustering, CIFAR-10, CIFAR-10-C, Mini-ImageNet, gradient magnitude, communication cost, FedACC, machine learning