Federated-Learning-Bias-CaseStudy

An in-depth case study analyzing bias in federated learning arising from label and sampling feature heterogeneity across edge devices. This repository provides the code to reproduce our experiments on CIFAR-10 and Cholec80, including data partitioning, FL training with normalization variants, evaluation, and visualization.

---

Table of Contents

1. Introduction
2. Installation & Prerequisites
3. Data Preparation
4. Configuration
5. Usage
   • Training
   • Evaluation
   • Visualization
6. Experiments & Results
7. Citation

---

Introduction

Federated Learning (FL) enables collaborative model training across edge devices without sharing raw data. However, heterogeneous sensors introduce two key sources of bias:

• Label heterogeneity: uneven class distributions across clients.
• Sampling feature heterogeneity: differences in feature representations (e.g., noise levels) due to diverse device quality.

Our empirical case study on CIFAR-10 and a surgical task dataset (Cholec80) shows that while normalization methods (BatchNorm, GroupNorm, InstanceNorm, LayerNorm) can improve overall performance, none eliminate per-client bias induced by sampling feature heterogeneity.

Contributions

1. Empirical analysis of label vs. sampling feature heterogeneity impact on per-client bias in FL.
2. Evaluation of state-of-the-art normalization techniques under both heterogeneity types.
3. Insights into performance–fairness–resource trade-offs, motivating new bias-mitigation strategies for FL.

---

Installation & Prerequisites

1. Clone the repository

   git clone https://github.com/emtechlab/federated-learning-biases.git
   cd federated-learning-biases