Confidential Computing: Protecting Medical Data While It's In Use
— by Vivax
Encryption protects data at rest and in transit — but to compute on it, software has traditionally had to decrypt it in memory, where it is exposed.
Data has three states: at rest (stored), in transit (moving across a network), and in use (loaded into memory for processing). We have long encrypted the first two. The third has been the weak point: to actually compute on data, a CPU normally decrypts it into memory, where a compromised operating system, hypervisor, or cloud administrator could read it. Confidential computing exists to protect data in use.
It does this with a Trusted Execution Environment (TEE) — a hardware-isolated region of a processor where code and data are protected from everything outside it, including privileged system software. The Confidential Computing Consortium, a Linux Foundation project, stewards the standards. A key feature is remote attestation: the hardware can cryptographically prove to a remote party exactly what code is running inside the enclave before any secret is sent to it.
The hardware options have matured quickly. Intel offers SGX (application-level enclaves) and TDX (full confidential virtual machines); AMD provides SEV-SNP for confidential VMs; and ARM's Confidential Compute Architecture (CCA) brings the model to its ecosystem. Critically for AI, confidential computing has reached the GPU: NVIDIA's H100 and successors can run accelerated workloads inside a TEE, so even large models can be trained or served on encrypted data with hardware-verified isolation.
For grounded clinical AI, this is foundational. Confidential computing lets sensitive patient data and proprietary models be processed in the cloud or across institutions while remaining encrypted and attestable end to end. It complements differential privacy and federated learning: where those bound what a model reveals and keep data local, TEEs protect the computation itself — a defense-in-depth approach to handling protected health information.