SplitFedZip: Learned Compression for Data Transfer Reduction in Split-Federated Learning

TL;DR

SplitFedZip targets the communication bottlenecks in SplitFed learning by introducing end-to-end rate-distortion optimized learned compression of both features and gradients. It deploys a three-part Split U-Net with two split points and trains two codecs (AE and Cheng_AT) within a federated framework, optimizing the loss $L = L_r + \lambda (L_{Dice} + L_{mse})$ to trade off bit-rate and task accuracy. Across two medical segmentation datasets, SplitFedZip achieves data-transfer reductions of several orders of magnitude while preserving or slightly improving final segmentation accuracy, with Cheng_AT providing stronger RD gains than the AE codec; when compared to prior SplitFed compression methods, SplitFedZip delivers substantially better data-efficiency with comparable or superior accuracy. This work enables scalable, privacy-preserving SplitFed deployments in healthcare by dramatically reducing communication overhead without sacrificing model quality.

Abstract

Federated Learning (FL) enables multiple clients to train a collaborative model without sharing their local data. Split Learning (SL) allows a model to be trained in a split manner across different locations. Split-Federated (SplitFed) learning is a more recent approach that combines the strengths of FL and SL. SplitFed minimizes the computational burden of FL by balancing computation across clients and servers, while still preserving data privacy. This makes it an ideal learning framework across various domains, especially in healthcare, where data privacy is of utmost importance. However, SplitFed networks encounter numerous communication challenges, such as latency, bandwidth constraints, synchronization overhead, and a large amount of data that needs to be transferred during the learning process. In this paper, we propose SplitFedZip -- a novel method that employs learned compression to reduce data transfer in SplitFed learning. Through experiments on medical image segmentation, we show that learned compression can provide a significant data communication reduction in SplitFed learning, while maintaining the accuracy of the final trained model. The implementation is available at: \url{https://github.com/ChamaniS/SplitFedZip}.

Figures (6)

• Figure 1: Proposed SplitFedZip architecture.
• Figure 2: R-A curves for the Blastocyst dataset.
• Figure 3: R-A curves for the HAM10K dataset.
• Figure 4: CR vs. $\lambda$ for the HAM10K dataset.
• Figure 5: Analysis of BPP during SplitFed training for the blastocyst dataset.