PrivFED -- A Framework for Privacy-Preserving Federated Learning in Enhanced Breast Cancer Diagnosis

TL;DR

This paper presents PrivFED, a privacy-preserving federated learning framework for breast cancer diagnosis that enhances data security, handles data imbalance, and achieves high accuracy using advanced techniques like SMOTE, isolation forests, PCA, and Catboost.

Contribution

It introduces a novel federated learning framework with integrated data balancing, outlier detection, and feature selection tailored for healthcare diagnostics.

Findings

Achieved 99.95% accuracy on edge devices

Demonstrated robustness against outliers and data imbalance

Validated effectiveness on Wisconsin breast cancer dataset

Abstract

In the day-to-day operations of healthcare institutions, a multitude of Personally Identifiable Information (PII) data exchanges occur, exposing the data to a spectrum of cybersecurity threats. This study introduces a federated learning framework, trained on the Wisconsin dataset, to mitigate challenges such as data scarcity and imbalance. Techniques like the Synthetic Minority Over-sampling Technique (SMOTE) are incorporated to bolster robustness, while isolation forests are employed to fortify the model against outliers. Catboost serves as the classification tool across all devices. The identification of optimal features for heightened accuracy is pursued through Principal Component Analysis (PCA),accentuating the significance of hyperparameter tuning, as underscored in a comparative analysis. The model exhibits an average accuracy of 99.95% on edge devices and 98% on the central server.