Equitable Federated Learning for Skin Cancer Detection: A Fairness-Driven Approach

Skin cancer is a prevalent malignancy where early detection is vital for improving patient outcomes. Traditional deep learning methods often rely on centralized data, raising concerns over privacy and security. Federated learning (FL) offers a decentralized alternative, enabling model training across multiple institutions without sharing sensitive data. However, the performance of existing FL approaches depends heavily on balanced data distribution, which is often unrealistic in realworld scenarios. In this study, we propose a fairness-aware FL framework for skin cancer detection, leveraging diverse and heterogeneous dermatological image datasets from multiple healthcare institutions. Unfair or biased client models can negatively impact global model convergence and overall performance. To address this, our framework incorporates a global validation dataset to evaluate each client’s local model during every training round. Specifically, the server evaluates the performance of client-submitted models against the global dataset and enforces a predefined performance threshold. Only client models that surpass this threshold are incorporated into the global model aggregation process, ensuring that only high-performing and equitable models contribute to the collective learning. The proposed framework is evaluated against baseline FL approaches using critical performance metrics, including accuracy, sensitivity, and specificity. Empirical results demonstrate that the fairnessaware FL framework surpasses traditional methods, achieving a maximum accuracy of 85.29% and enhanced robustness in the detection of skin cancer. In our approach, the overall performance improvement ranges from 5% to 8%, depending on the degree of unfairness and the number of clients involved.