BFCL: a secure decentralized framework for chest x-ray image classification using blockchain federated contrastive learning

Secure and privacy-preserving collaboration among healthcare institutions is essential for advancing medical diagnosis. We propose a blockchain-enabled federated supervised contrastive learning framework for the multiclass classification of chest X-ray images. Unlike existing blockchain-based federated learning approaches, our method introduces a hybrid loss function that aligns local and global feature representations, improving generalization under heterogeneous and non-IID data. The blockchain layer goes beyond secure aggregation by incorporating a Proof-of-Stake consensus mechanism for validating client updates, thereby enhancing robustness against unreliable or adversarial contributions. To further strengthen security, we assess resilience against poisoning attacks, with a focus on label-flipping scenarios. Experimental results show that our framework significantly outperforms conventional federated learning in both classification accuracy and resistance to attacks, while also ensuring transparency, privacy preservation, and reduced communication overhead. This dual integration of contrastive alignment and consensus-driven validation provides a novel paradigm for medical federated learning.