Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Smart Split-Federated Learning over Noisy Channels for Embryo Image Segmentation

Zahra Hafezi Kafshgari, Ivan V. Bajic, Parvaneh Saeedi

TL;DR

The paper addresses privacy-preserving distributed learning for medical image analysis by examining SplitFed under noisy communication channels. It introduces Smart SplitFed, a loss- and data-aware weighted averaging scheme that uses per-client loss statistics to adjust contributions during aggregation. Applied to a SplitFed U‑Net for embryo image segmentation, Smart SplitFed achieves two orders of magnitude greater tolerance to channel noise than naive averaging while preserving final accuracy. This approach enhances the practicality of SplitFed in real-world healthcare deployments where communication links are imperfect.

Abstract

Split-Federated (SplitFed) learning is an extension of federated learning that places minimal requirements on the clients computing infrastructure, since only a small portion of the overall model is deployed on the clients hardware. In SplitFed learning, feature values, gradient updates, and model updates are transferred across communication channels. In this paper, we study the effects of noise in the communication channels on the learning process and the quality of the final model. We propose a smart averaging strategy for SplitFed learning with the goal of improving resilience against channel noise. Experiments on a segmentation model for embryo images shows that the proposed smart averaging strategy is able to tolerate two orders of magnitude stronger noise in the communication channels compared to conventional averaging, while still maintaining the accuracy of the final model.

Smart Split-Federated Learning over Noisy Channels for Embryo Image Segmentation

TL;DR

The paper addresses privacy-preserving distributed learning for medical image analysis by examining SplitFed under noisy communication channels. It introduces Smart SplitFed, a loss- and data-aware weighted averaging scheme that uses per-client loss statistics to adjust contributions during aggregation. Applied to a SplitFed U‑Net for embryo image segmentation, Smart SplitFed achieves two orders of magnitude greater tolerance to channel noise than naive averaging while preserving final accuracy. This approach enhances the practicality of SplitFed in real-world healthcare deployments where communication links are imperfect.

Abstract

Split-Federated (SplitFed) learning is an extension of federated learning that places minimal requirements on the clients computing infrastructure, since only a small portion of the overall model is deployed on the clients hardware. In SplitFed learning, feature values, gradient updates, and model updates are transferred across communication channels. In this paper, we study the effects of noise in the communication channels on the learning process and the quality of the final model. We propose a smart averaging strategy for SplitFed learning with the goal of improving resilience against channel noise. Experiments on a segmentation model for embryo images shows that the proposed smart averaging strategy is able to tolerate two orders of magnitude stronger noise in the communication channels compared to conventional averaging, while still maintaining the accuracy of the final model.
Paper Structure (10 sections, 7 equations, 3 figures, 1 table)

This paper contains 10 sections, 7 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related works
  3. Proposed Method
  4. Segmentation model
  5. Training procedure
  6. Averaging
  7. Experiments
  8. Experimental setup
  9. Results
  10. Conclusions

Figures (3)

  • Figure 1: SplitFed U-Net used in our work.
  • Figure 2: A comparison of Naive SplitFed, SplitFedAVG, and Smart SplitFed at $\sigma_{\text{noise}} = 6\cdot 10^{-4}$. Top row: training loss at each client. Middle row: clients' weights used in averaging. Bottom row: test loss at each global epoch.
  • Figure 3: (a) Input image, (b) ground truth segmentation, and predictions made by models trained at $\sigma_{\text{noise}} = 6\cdot 10^{-4}$ using (c) Naive SplitFed, (d) SplitFedAVG, and (e) Smart SplitFed.