Proposal: Verifiable Confidential Execution for MemPrivacy

[Jasonmils]

Problem

MemPrivacy currently protects sensitive data primarily through local detection and pseudonymization. However:

1. Privacy detection is probabilistic and may miss sensitive spans.
2. Users still need to trust the remote service not to inspect or retain plaintext that passes through.
3. Open-source code alone cannot prove that the same code is running in production.

Proposal

Add an optional Confidential Execution Mode based on TEE + Remote Attestation.

Client
  │
  ├─ Verify remote attestation
  ├─ Verify expected workload measurement
  │
  └─ Encrypt request with attested public key
              ↓
       ┌─────────────┐
       │     TEE     │
       │ MemPrivacy  │
       │ Processing  │
       └─────────────┘
              ↓
       Encrypted result

The client should only release sensitive data or encryption keys after verifying that:

• the workload is running inside a supported TEE;
• its code measurement matches an approved build;
• debug mode is disabled.

Why

This changes MemPrivacy's security model from:

"Sensitive information is unlikely to leave the client if correctly detected."

to:

"Even if sensitive information reaches the server, the infrastructure operator cannot directly access it."

A minimal first implementation could target one backend such as AWS Nitro Enclaves, keeping the existing local redaction pipeline unchanged.

This would make privacy defense-in-depth rather than dependent solely on detection accuracy.