Confidential Computing

Traditional encryption protects data at rest (storage) and in transit (network), but data must be decrypted in system memory to be processed. Confidential Computing closes this final gap by protecting data in use.

• Encrypted RAM: The CPU encrypts data as it moves between the processor cache and main memory, rendering a physical memory dump useless.
• Protection from Insiders: System administrators, cloud operators, and anyone with root access to the host machine are cryptographically blocked from viewing the plaintext data being processed.

The TEE guarantees that the code executing inside the enclave has not been tampered with, either before launch or during runtime.

• Launch Measurement: The hardware measures the exact code and initial data loaded into the enclave, creating a unique cryptographic identity.
• Runtime Protection: The enclave's memory pages are protected from external modification. Any attempt to alter the executing code or its data by the host OS or a DMA attack is blocked by the hardware.

The security of the entire TEE rests on a Hardware Root of Trust—a set of immutable, cryptographically verifiable keys and logic physically embedded in the processor silicon during manufacturing.

• Key Derivation: All enclave-specific encryption keys are derived from this unique, un-extractable root key.
• Chain of Trust: This root anchors a verifiable chain of trust that extends from the silicon up to the application, ensuring every component in the stack is authenticated.

A cryptographic verification process that proves to a remote party that a specific workload is running unmodified inside a genuine TEE on trusted hardware.

• Goal: Establish trust before releasing secrets or data to an enclave.
• Process: The TEE generates a hardware-signed report (quote) of its initial state and identity, verified against the manufacturer's endorsement key.
• Critical for: Multi-party federated learning where data owners must verify the aggregation environment.

A cryptographic protocol enabling multiple parties to jointly compute a function over their private inputs while keeping those inputs mutually secret. Unlike TEEs, SMPC relies on mathematics rather than hardware trust.

• Trade-off: Higher computational/communication overhead vs. no hardware root of trust required.
• Synergy: Often combined with TEEs in hybrid models for defense-in-depth.
• Primitives: Garbled circuits, secret sharing, oblivious transfer.

A cryptographic scheme enabling computation directly on encrypted data without requiring decryption. The result remains encrypted and matches the computation on plaintext.

• Types: Partially (PHE), Somewhat (SHE), and Fully Homomorphic Encryption (FHE).
• Confidential Computing Contrast: HE protects data during computation via mathematics; TEEs protect it via hardware isolation.
• Current Limitation: Significant computational overhead makes FHE impractical for large-scale deep learning training.

The hardware mechanism that transparently encrypts and decrypts data moving between the processor cache and main memory (RAM) within a TEE.

• Purpose: Defends against physical bus snooping, cold boot attacks, and DRAM interposers.
• Implementation: A dedicated on-die memory encryption engine with per-enclave or per-VM keys.
• Performance: Modern AES-XTS engines operate at near-line rate with minimal latency impact.

The application of Confidential Computing to decentralized model training to protect against gradient leakage and malicious aggregation servers.

• Architecture: The central aggregation server runs inside a TEE, preventing the cloud operator from inspecting individual client model updates.
• Attestation Flow: Clients verify the aggregator's TEE identity via remote attestation before transmitting local gradients.
• Complementary Defense: Used alongside Differential Privacy and Secure Aggregation for a multi-layered privacy posture.