Skin Lesion Classification using a Decentralized Peer-to-Peer Federated Learning Framework

Early identification of skin cancer is crucial to improve patient outcomes. However, traditional centralized deep learning poses privacy risks and logistical challenges. Federated learning offers a solution that enables collaborative model training on decentralized devices without data exchange. This paper presents a peer-to-peer federated learning approach using the HAM10000 dataset of dermatoscopic images. In our method, each client uses the same model architecture and iteratively updates model weights based on the marginal contribution of local data. This ensures robust and reflective updates and improves learning with local data characteristics. This approach preserves data privacy, reduces logistical challenges, and improves model accuracy by leveraging diverse datasets. Increasing the number of clients correlates with better accuracy and generalizability. With four clients and a VGG11 model enhanced with a self-attention layer, we have achieved 92.84% classification accuracy, 93.02% precision, 92.84% recall, and a 92.89% F1-score.