Fed-Focal Loss for Imbalanced Data Classification in Federated Learning

Jan 8, 2021·
Dipankar Sarkar
,
Ankur Narang
,
Sumit Rai
· 1 min read
Abstract
This paper extends the Focal Loss function used in image detectors to Federated Learning along with a tunable sampling framework for solving the class imbalance problem. The approach reshapes cross-entropy loss to down-weight the loss assigned to well-classified examples, following focal loss principles. Additionally, it leverages a tunable sampling framework to account for selective client model contributions on the central server, improving detector focus during training and enhancing robustness.
Date
Jan 8, 2021 10:30 AM — 10:45 AM
Event
Location

Blue Wing-North 4 (VirtualChair Gathertown)

Presentation Details

This talk is part of Technical Talks Session 2 at the FL-IJCAI'20 workshop. The presentation will cover:

  • Extension of Focal Loss to Federated Learning context
  • Implementation of Fed-Focal loss function
  • Tunable sampling framework for client selection
  • Experimental results across multiple datasets
  • Impact on training stability and model robustness

Reviews

The paper received positive reviews highlighting:

  • Novel application of Focal Loss in Federated Learning context
  • Comprehensive experimental evaluation
  • Promising results in terms of accuracy and robustness
  • Well-written presentation of the methodology

Workshop Information

The International Workshop on Federated Learning for User Privacy and Data Confidentiality (FL-IJCAI'20) focuses on machine learning systems adhering to privacy-preserving and security principles. The workshop provides a forum to discuss open problems and share ground-breaking work in secure and privacy-preserving compliant machine learning.