To address the challenges outlined earlier, we introduce an integrated framework for Biomedical Decentralized Science (BioDeSci) that combines:
Secure data exchange
Private model collaboration
Verifiable computation
This end-to-end system is structured around three coordinated levels: Data, Model, and Evaluation, as illustrated below.
Core Components
Data Level
A decentralized biomedical data marketplace allows hospitals, pharmaceutical firms, and academic labs to share encrypted datasets. Tokenized metadata and privacy-preserving smart contracts govern access. The system supports secure upload, fine-grained access control, and verifiable provenance—without ever disclosing raw patient information.
Model Level
Rexis applies Equivariant Encryption (EE) to enable decentralized fine-tuning of large biomedical models (e.g., transformers). Each institution performs encrypted training locally. Gradients are securely aggregated into a global model, without exposing raw data, intermediate representations, or model internals.
Evaluation Level
EE also enables encrypted inference. Outputs are verified via on-chain commitments, such as hash references and lightweight cryptographic proofs. External auditors can confirm model integrity and output correctness—without decrypting private data.
Together, these components form a scalable, verifiable, and privacy-preserving framework for biomedical AI. The system:
Preserves confidentiality
Guarantees computational integrity
Facilitates reproducibility across institutions
These properties are essential for regulatory compliance and scientific credibility in real-world biomedical deployments.
Overview of Data, Model, and Evaluation Levels. Our end-to-end pipeline operates across three coordinated layers: (1) Data Level: Hospitals, research labs, and pharmaceutical companies contribute encrypted biomedical datasets to a decentralized data marketplace, with blockchain-backed tokenization and access control; (2) Model Level: Equivariant Encryption (EE) enables privacy-preserving fine-tuning and secure aggregation of encrypted models across institutions without exposing raw data; (3) Evaluation Level: Encrypted inference is performed with verifiable integrity via lightweight cryptographic proofs and on-chain audit records, ensuring correctness and privacy.
