FedCIAL: Federated Color-Invariant Adversarial Learning for Enhancing Fairness and Performance in Skin Lesion Classification

Federated learning (FL) enables healthcare institutions to collaboratively train a global model without sharing patient data. However, this approach can still introduce bias into the learning process. In the context of skin lesion classification, bias is a significant challenge due to the diversity in skin tone representation. Models trained on imbalanced datasets that underrepresent darker skin tones may exhibit reduced accuracy and reliability for those population. In this work, we propose Federated Color-Invariant Adversarial Learning (FedCIAL), a novel approach that leverages known color distribution shifts to generate target samples. This allows us to train a color-invariant feature extractor using domain adaptation techniques without requiring any shared data. Experimental results on the Fitzpatrick17k dataset show that FedCIAL outperforms the state-of-the-art model FeSViBS, achieving an average accuracy of 0.7754, compared to 0.7666 for the baseline, with a statistically significant improvement (p = 0.044). Additionally, FedCIAL improves model fairness, reducing the standard deviation across clients to 0.044, compared to 0.053 for the baseline. These findings demonstrate that FedCIAL enhances performance and offers a promising solution for fairer federated learning models in medical image analysis.