MUFFLe: Efficient Model Update Compression via Generalized Deduplication for Federated Learning

2026-06-12

AI Summary

This work addresses the high communication overhead associated with uploading model updates in federated learning by introducing generalized deduplication into the FedAvg framework for the first time. The proposed method employs a fixed-rate, variable-quantity compression mechanism that identifies and exploits repetitive patterns within model update vectors to achieve efficient communication. Experimental results on the IID MNIST dataset demonstrate that the approach attains an accuracy of 92.93% with only 38 MB of uplink communication, substantially outperforming 8-bit quantization (75 MB), Top-k sparsification (86 MB), and uncompressed FedAvg (310 MB). These findings confirm the method’s effectiveness and superiority in significantly reducing communication costs while maintaining high model accuracy.

Abstract

Federated learning is well suited to edge environments but is often limited by the uplink cost of transmitting model updates. This Work-in-Progress paper presents MUFFLe, a communication-efficient update compression scheme that integrates generalized deduplication (GD) into the FedAvg pipeline. MUFFLe deduplicates repeated patterns across the update vector, yielding a fixed-rate, variable-count compression scheme. Preliminary experiments on IID MNIST with 20 clients show that MUFFLe reaches the target accuracy of $92.93\%$ with 38~MB cumulative uplink communication, compared with 75~MB for 8-bit quantization, 86~MB for Top-$k$ sparsification, and 310~MB for uncompressed FedAvg. These results demonstrate the feasibility of applying GD to communication-efficient federated learning.

Problem

• federated learning
• communication cost
• model update compression
• uplink bandwidth

Innovation

• generalized deduplication
• federated learning
• model update compression
• communication efficiency
• FedAvg