Model Encryption

A layered cryptographic strategy where a Data Encryption Key (DEK) encrypts the model artifact, and a Key Encryption Key (KEK) wraps the DEK. This enables efficient key rotation without re-encrypting large model files.

• DEK: Symmetric key (e.g., AES-256-GCM) for bulk model encryption
• KEK: Asymmetric or symmetric key managed by a KMS
• Benefit: Revoke access instantly by deleting the KEK

Decryption keys are released only after a Trusted Execution Environment (TEE) or Secure Element proves its identity and integrity via remote attestation. The key never leaves the secure hardware boundary.

• Attestation verifies firmware hash and security configuration
• Keys are bound to a specific device identity (e.g., Physically Unclonable Function)
• Prevents key extraction from compromised host operating systems

Model weights remain encrypted in persistent storage and are only decrypted in-memory within a protected enclave immediately before execution. Cleartext weights are never written to disk.

• Minimizes the window of exposure for plaintext parameters
• Memory pages can be pinned to prevent swapping to unencrypted storage
• Integrates with bus encryption to protect data in transit between CPU and RAM

Before decryption, the runtime verifies the model artifact's cryptographic hash and digital signature to ensure it hasn't been tampered with or corrupted. This establishes a chain of trust from the model author.

• Uses SHA-384 or SHA-512 for collision resistance
• Signatures verified against a hardware-fused public key or certificate chain
• Detects both malicious tampering and storage bit-rot

Active defense mechanisms that trigger upon detecting unauthorized access attempts. Responses include zeroization of keys and model weights, halting inference, and incrementing a tamper-evident counter.

• Side-channel attack detection (timing, power analysis)
• JTAG/SWD debug port disablement in production fuses
• Mesh sensors in hardware security modules detect physical intrusion

Encrypted model artifacts are distributed through signed update packages over authenticated channels. Each device receives a uniquely wrapped DEK, enabling per-device access control and audit trails.

• Mutual TLS secures the transport layer
• Device-specific key derivation prevents one compromised key from exposing all deployments
• Supports air-gapped transfer via encrypted removable media

A cryptographic method enabling computation directly on ciphertext. The model processes encrypted inputs and produces encrypted outputs, ensuring the runtime environment never sees plaintext data. Fully Homomorphic Encryption (FHE) supports arbitrary computations but incurs high latency; Partially Homomorphic Encryption (PHE) is faster but limited to specific operations. Key schemes include CKKS for approximate arithmetic and TFHE for fast bootstrapping.

A hardware-enforced secure area within a main processor that isolates code and data from the host OS, hypervisor, and other applications. Intel SGX and AMD SEV create encrypted memory regions where model weights are decrypted and inference runs. TEEs provide confidential computing guarantees: even a compromised kernel cannot read enclave memory. Remote attestation cryptographically verifies the enclave's identity before key release.

Partitions a neural network's computational graph and parameters across multiple isolated devices or secure enclaves. No single node holds the complete model, raising the bar for extraction attacks. Techniques include: horizontal sharding (splitting layers across devices), vertical sharding (partitioning neurons within layers), and secret sharing schemes where weights are mathematically split into meaningless shares. Combines naturally with SMPC for secure distributed inference.

Defenses against attacks that infer model secrets from physical information leakage rather than cryptographic weaknesses. Timing attacks measure execution latency to deduce weights; mitigations use constant-time operations. Power analysis monitors current draw; defenses employ power-balancing circuits. Electromagnetic emanations can reveal computation patterns; shielding and noise generators mask the signal. Cache-based side channels are addressed through cache coloring and prefetch randomization.

Cryptographic implementations designed to protect secret keys even when an attacker has full visibility into the execution environment and can observe or modify memory and internal operations. Applied to model encryption, white-box techniques embed the decryption key into a heavily obfuscated algorithm that is mathematically equivalent but resistant to key extraction. Common approaches include networked encoding, state-dependent transformations, and asymmetric backdoors that make reverse engineering computationally prohibitive.

A cryptographic protocol distributing computation across multiple parties who jointly compute an inference result without revealing their private input shares. For model encryption, SMPC enables threshold decryption: the model key is split into shares, and a quorum of parties must collaborate to decrypt and run inference. Even if an attacker compromises some nodes, the model remains protected. Common protocols include SPDZ for arithmetic circuits and Garbled Circuits for boolean operations.