NOVA + Phala TEE Demo: Verifiable Federated Learning Datasets

Phala's TEEs ensure confidential training, but lack persistent storage. NOVA complements TEEs by providing secure vaults for datasets pre/post-TEE, with revokable group membership. This enables verifiable, multi-party workflows: upload shared data, TEE fine-tunes, store models back—auditable via on-chain metadata.

Use Case

In healthcare (e.g., Phala's success story: https://phala.com/success-stories/healthcare-research), hospitals share encrypted records for federated learning without exposure.

In the following demo, Hospital A uploads encrypted data to NOVA, Hospital B retrieves and process the data through Phala's TEE, it uploads the output data to NOVA. Hospital A retrieves Hospital B's output data, process it, and uploads the final output to NOVA.

Other Scenarios

• Finance: Secure Portfolio Sharing Before TEE Risk Analysis (https://phala.com/success-stories/financial-services): NOVA's group-keyed vaults store encrypted portfolios; TEE loads for private risk modeling (e.g., fraud detection), outputs back without leaks—ensures data sovereignty in multi-bank collaborations.

• Law: Confidential Document Vaults Before TEE Review (https://phala.com/posts/confidential-ai-for-law-firms): Upload sensitive legal docs to NOVA; TEE processes (e.g., AI contract analysis) without visibility; revoke access post-review—complements Phala's privacy for e-discovery.

• Decentralized AI: Persistent Agent State for Multi-Chain Agents (https://phala.com/success-stories/decentralized-ai): NOVA persists encrypted agent states/models between TEE sessions; enables multi-chain autonomy (e.g., cross-NEAR/Solana inference) with attestation chaining.

How NOVA Complements Private ML SDK

Phala focuses on runtime privacy ("no storage, no logs"); NOVA adds decentralized persistence: encrypt/upload via composite_upload, retrieve into TEE, output back. Focus: Data pipeline security without reinventing Phala's compute. NOVA handles pre-enclave sourcing (group auth) and post-enclave auditing (on-chain hashes/CIDs), extending Phala's quotes for full-lifecycle verifiability.

Running the Demos

• Set .env (NEAR_PRIVATE_KEY, etc.).

• Rust: cargo run --bin tee_federated_learning.

• JS: ts-node demos/tee_federated_learning.ts.

Expected output:

Primary Account ID (Hospital A): <hospital_A.testnet>
Secondary Account ID (Hospital B): <hospital_B.testnet>
Hospital A uploaded to NOVA: CID QmW9oCqbrdMF8cKuYd14cwT3SszMWXRnaJ4aTZXKG4b1QA
Waiting for IPFS pin to propagate...
Hospital B output stored: CID QmSxuYEMhJpCm2wxC9zRQrbvaJN2irGtNdENCxRzs6J39C
Waiting for IPFS pin to propagate...
Hospital A final output stored: CID QmP5TiZYHByf1DDtXxVe5KjHj8r5Sp2E2MzLCMAxfH3Job