Trusted Execution Environment (TEE)

The TEE creates a secure enclave that is physically isolated at the CPU level. Code and data inside the enclave are protected from all software outside it, including the OS kernel, hypervisor, and DMA attacks. This is achieved through memory encryption engines that automatically encrypt and decrypt data as it moves between the processor cache and external RAM, ensuring plaintext data is never exposed on the memory bus.

A cryptographic mechanism that allows a remote party to verify the exact identity and integrity of the software running inside a TEE. The processor generates a signed attestation report containing a cryptographic hash of the enclave's initial state and code. This proves to a client that they are communicating with a genuine, untampered application running on a legitimate TEE-capable platform, not a simulator or compromised host.

The TEE guarantees that data in use—the most vulnerable state in the data lifecycle—remains encrypted and inaccessible. Key protections include:

• Confidentiality: The host OS, hypervisor, or a physical attacker with a logic probe cannot read plaintext data inside the enclave.
• Integrity: Any attempt to modify enclave memory is detected and blocked, preventing rollback or tampering attacks.
• Sealing: Data can be encrypted and persisted to disk, bound to a specific enclave identity, ensuring only that exact code on that specific CPU can decrypt it later.

The TCB is the set of all hardware, firmware, and software components critical to the system's security. A TEE dramatically reduces the TCB by excluding the entire operating system, device drivers, and hypervisor from the trust boundary. The only trusted components are the CPU silicon itself, the enclave code, and the thin security monitor. This reduction minimizes the attack surface and simplifies formal verification.

The TEE manages enclave thread scheduling securely, ensuring that when the CPU context-switches between an enclave and the untrusted OS, all enclave register state is automatically saved and encrypted. The Asynchronous Enclave Exit (AEX) mechanism handles interrupts and exceptions by flushing sensitive state to protected memory before the OS handler executes, preventing the OS from observing or manipulating the enclave's execution state.

Modern TEE implementations incorporate hardware and microcode defenses against cache-timing and speculative execution attacks. Protections include:

• Cache partitioning to prevent the OS from priming caches to observe enclave access patterns.
• Speculative execution barriers that prevent transient instructions from leaking secrets.
• Constant-time cryptographic libraries that eliminate data-dependent timing variations within enclave code.
• Page-fault attack mitigation by restricting OS control over enclave page tables.

A cryptographic mechanism that allows a relying party to verify that a specific workload is running inside a genuine Trusted Execution Environment on trusted hardware. The TEE generates a signed report containing measurements of its internal state.

• Evidence: A signed token proving the TEE's identity and code hash
• Verifier Service: Validates the attestation evidence against trusted reference values
• Secret Provisioning: Secrets are released only after successful attestation

A private region of memory within a TEE that is isolated from the main operating system. Even a compromised kernel cannot read enclave memory. Intel SGX and AMD SEV are prominent implementations.

• Enclave Page Cache (EPC): Encrypted memory region holding enclave code and data
• Sealing: Encrypts enclave secrets for persistent storage outside the TEE
• Thread Control Structure (TCS): Manages logical processor entry and exit

A hardware component integrated into the memory controller that transparently encrypts and decrypts data moving between the CPU and RAM. This is the foundational mechanism enabling Trusted Execution Environments.

• AMD SME: Encrypts entire system memory with a single key
• Intel TME/MKTME: Total Memory Encryption with optional multi-key support
• AES-XTS: The standard cipher mode used for memory encryption at line speed

Combines TEEs with decentralized training to protect both model gradients and aggregation logic. A TEE-based secure aggregator ensures the central server cannot inspect individual client updates.

• Gradient Leakage Defense: TEE prevents the aggregator from logging raw gradients
• Malicious Node Mitigation: Attestation verifies client integrity before participation
• Differential Privacy Integration: Noise can be injected inside the TEE for provable guarantees

Methods to protect the intellectual property of neural network weights and architecture. When combined with a Trusted Execution Environment, the model is never exposed in plaintext to the host.

• Weight Encryption: Decryption occurs only inside the TEE during inference
• Operator Fusion: Obfuscates the model graph to resist reverse engineering
• Code Morphing: Runtime transformation of enclave code to deter static analysis