arXiv
Catfedavg: Optimising communication-efficiency and classification accuracy in federated learning
Dipankar Sarkar , S Rai , A Narang
We propose CatFedAvg, a novel approach to federated learning that optimizes both communication efficiency and classification accuracy. Our method introduces categorical federated averaging, which strategically aggregates model updates while reducing communication overhead. Through extensive experimentation, we demonstrate that CatFedAvg achieves comparable accuracy to traditional federated learning approaches while significantly reducing communication costs.
Abstract
Federated Learning (FL) enables distributed model training across multiple devices while preserving data privacy. However, communication efficiency remains a significant challenge in FL deployments. This paper introduces CatFedAvg, a novel federated learning algorithm that optimizes the trade-off between communication costs and model performance. By leveraging categorical aggregation strategies, our approach reduces the communication overhead while maintaining competitive classification accuracy. We evaluate CatFedAvg on various datasets and demonstrate its effectiveness in reducing communication rounds while achieving comparable or better performance than baseline FL methods.
The full paper is available on arXiv.
Frequently Asked Questions
What is CatFedAvg?
CatFedAvg is a federated learning approach that optimises both communication efficiency and classification accuracy. It introduces categorical federated averaging, which strategically aggregates model updates while reducing communication overhead. Published November 2020 on arXiv (2011.07229). 4 citations as of 2026.
How does CatFedAvg reduce communication costs?
CatFedAvg introduces categorical federated averaging: model updates are grouped into categories, and only the most informative categories are transmitted. This reduces the communication overhead per round while maintaining classification accuracy comparable to standard FedAvg. The approach is particularly valuable in cross-device federated learning settings where communication is the bottleneck.
How does CatFedAvg relate to Fed-Focal Loss?
Fed-Focal Loss (93 citations, FL-IJCAI 2020) and CatFedAvg (4 citations) are companion papers from the same November 2020 publication cycle. Both address practical challenges in real-world federated learning: Fed-Focal Loss addresses class imbalance, CatFedAvg addresses communication efficiency. They share two co-authors (Ankur Narang and Sumit Rai).