model verification

**Model Verification** in the context of AI security is the **process of verifying that a deployed model has not been tampered with, corrupted, or replaced** — ensuring model integrity by checking that the model in production matches the validated, approved version. **Verification Methods** - **Hash Verification**: Compute a cryptographic hash of model weights and compare to the approved hash. - **Behavioral Probes**: Send known test inputs and verify expected outputs match the validated model. - **Weight Checksums**: Periodic checksum of weight files detects unauthorized modifications. - **TEE Verification**: Run inference in a Trusted Execution Environment (TEE) that verifies model integrity. **Why It Matters** - **Supply Chain**: Verify that a model received from a third party hasn't been trojaned or modified. - **Production Safety**: Ensure the model controlling fab equipment is the approved, validated version. - **Compliance**: Regulatory requirements may mandate model integrity verification in production. **Model Verification** is **trust but verify** — ensuring that the deployed model is exactly the model that was validated and approved.