Secure Enclave AI Training for Healthcare
Secure enclaves are the hardware foundation of private clinical AI training
A secure enclave — also known as a Trusted Execution Environment (TEE) — is a hardware-isolated region of CPU memory where code and data can execute without being observable or modifiable by the operating system, hypervisor, or any other process — even with root or physical access to the machine.
In a clinical AI context, the secure enclave creates a verifiable boundary around the training process: the hospital's IT administrator cannot inspect training data inside the enclave. The AI company's model code cannot exfiltrate data outside the enclave. The enclave's identity — its cryptographic measurement (MRENCLAVE) — is independently verifiable through Intel's Data Center Attestation Primitives (DCAP).
How Intel SGX protects clinical data during AI training
- Enclave creation: The Rapha Edge Core OS C launcher creates an SGX enclave in protected memory. The enclave is measured — every byte of code loaded into it is hashed to produce a unique MRENCLAVE value.
- Enclave attestation: The Rapha Network Orchestration Hub verifies the enclave's identity through Intel DCAP: the MRENCLAVE must match the expected value for the Rapha training runtime, the MRSIGNER must match the Rapha signing key, and the Intel root CA certificate chain must validate.
- Data loading: Clinical data is read from hospital storage into the enclave's encrypted memory. The data is decrypted only inside the enclave. The operating system sees only encrypted memory pages.
- Model training: The AI model training executes entirely within the enclave. Memory access patterns are obfuscated. The CPU enforces that only enclave code can access enclave memory.
- Output extraction: Trained model weights are serialised within the enclave, hashed for the proof receipt, and transmitted to the researcher. No raw clinical data crosses the enclave boundary.
- Enclave destruction: After training completes, the enclave is torn down. All enclave memory pages are deallocated. The clinical data in memory is unrecoverable.
SGX vs TDX: which enclave technology for which workload?
Intel SGX (Software Guard Extensions)
- Protection scope: Application-level. Individual code regions within a process run inside the enclave.
- Memory limit: 512MB-1TB Enclave Page Cache (EPC), depending on processor generation.
- Best for: Targeted ML training workloads where the training script, hyperparameters, and data loader are known and bounded.
- Rapha Protocol support: Primary TEE. Production C enclave launcher. Gramine manifest for SGX execution.
Intel TDX (Trust Domain Extensions)
- Protection scope: VM-level. The entire virtual machine — OS, libraries, application — runs inside a hardware-isolated trust domain.
- Memory limit: Full VM memory allocation. Supports larger workloads than application-level SGX.
- Best for: Complex ML workflows requiring full OS access, custom libraries, or large memory footprints that exceed SGX enclave limits.
- Rapha Protocol support: Planned. TDX support roadmap for workloads requiring VM-level isolation.
Production-grade SGX execution requires: Linux host with Intel SGX/DCAP driver stack, configured Gramine manifest, expected MRENCLAVE/MRSIGNER values, Intel root CA PEM, and trusted attestor signing key. Non-SGX hardware returns ENOTSUP — no software fallback, no synthetic attestation, no path to settlement.