Federated & Private ML

ClusterGuardFL: Secure Cluster-Weighted Client Aggregation

Enhancing Federated Learning Through Secure Cluster-Weighted Client Aggregation

Federated learning that groups clients by how they behave, so a minority of bad actors cannot poison the shared model.

The problem

Clients in federated learning differ wildly in data and behaviour, and some may send corrupted or malicious updates.

The approach

Dissimilarity scores between local and global models drive k-means clustering; cluster size and a per-client reconciliation confidence feed a softmax that sets each client’s aggregation weight.

The outcome

Outperformed state-of-the-art aggregation under label-flipping and Gaussian-noise attacks on MNIST, Fashion-MNIST and CIFAR-10, holding accuracy with 20% malicious clients.

In detail

Proposed a cluster-weighted aggregation strategy for federated learning under heterogeneous client data. The framework uses dissimilarity measures, clustering and confidence-based reconciliation to identify groups of similar client updates and assign aggregation weights dynamically. The approach is intended to improve robustness, convergence and fairness when client data quality and behaviour vary significantly.

Built with

• Federated Learning
• Robustness
• Client Reliability

Source

• Enhancing Federated Learning Through Secure Cluster-Weighted Client Aggregation, arXiv:2503.22971, 2025

More in Federated & Private ML

• IRT-FedLoRA: Reliability-Aware Federated Fine-Tuning
• AR-FL: Anomaly-Resilient Federated Learning
• DP-FedMeta: Differentially Private Federated Meta-Learning

Next project: AR-FL: Anomaly-Resilient Federated Learning

By Kanishka Ranaweera, PhD, Machine Learning Engineer in Melbourne, Victoria, Australia. hello@kanishkabuilt.me