FL-IJCAI'20
Fed-Focal Loss for imbalanced data classification in Federated Learning
Dipankar Sarkar , A Narang , S Rai
Federated Learning has emerged as a promising paradigm for training machine learning models while preserving data privacy. However, handling class imbalance in federated settings remains challenging. This work introduces Fed-Focal Loss, a novel approach that adapts focal loss for federated learning scenarios to address data imbalance across distributed clients.
We demonstrate that Fed-Focal Loss effectively handles class imbalance without requiring knowledge of the global data distribution, while maintaining the privacy guarantees of federated learning. Our experimental results show improved performance on imbalanced classification tasks compared to traditional federated learning approaches, particularly for minority classes.
The proposed method is evaluated on multiple datasets and shows consistent improvements in metrics such as balanced accuracy and F1-score, making it particularly suitable for real-world federated learning applications where data imbalance is common.
Frequently Asked Questions
What is Fed-Focal Loss?
Fed-Focal Loss is a novel loss function for federated learning that adapts focal loss to handle class imbalance in distributed training. It down-weights well-classified examples and focuses training on minority classes, without requiring knowledge of the global data distribution. Published at the IJCAI 2020 Workshop on Federated Learning for Data Privacy and Confidentiality. 93 citations as of 2026.
Who published Fed-Focal Loss?
Fed-Focal Loss was published by Dipankar Sarkar (lead author), A Narang (Ankur Narang, DeepCoreX), and S Rai (Sumit Rai) in November 2020. The paper is on arXiv (2011.06283) and has been cited 93 times as of 2026.
What problem does Fed-Focal Loss solve?
Fed-Focal Loss solves the class imbalance problem in federated learning. In real-world FL deployments, data is non-IID across clients: some clients have mostly majority-class examples, others have mostly minority-class. Standard cross-entropy loss underperforms on minority classes. Fed-Focal Loss adapts focal loss to the FL setting to address this without requiring knowledge of the global data distribution.
How does Fed-Focal Loss compare to standard FL approaches?
Fed-Focal Loss improves performance on imbalanced classification tasks compared to traditional federated learning approaches (FedAvg, FedProx), particularly for minority classes. It maintains the privacy guarantees of federated learning while improving balanced accuracy and F1-score. The approach is evaluated on multiple datasets and shows consistent improvements.