What are the best practices for offline AI model security and tamper prevention?

Securing offline AI models requires multiple layers: model encryption at rest using AES-256 with hardware-backed key storage (TPM, Secure Enclave, ARM TrustZone), code signing for model files to detect tampering, obfuscation of model architecture through weight encryption and custom inference runtimes, and runtime integrity verification using hash chains. Implementation pattern: encrypt model weights, decrypt only into secure memory during inference, implement model watermarking for IP tracking, use hardware attestation to verify the execution environment before loading models. Prevent extraction attacks: disable debugging interfaces, implement anti-hooking mechanisms, use TEE (Trusted Execution Environment) for sensitive inference. For adversarial robustness, apply input validation, randomized smoothing, and anomaly detection on inputs. Regular security audits should test for model inversion attacks, membership inference attacks, and model stealing attempts.