Inference Verification in a Trusted Execution Environment (TEE)

We use inference verification to verify that text actually came from a model. This runs on an Nvidia H200 TEE, with the TEE code built and hosted by tinfoil.sh.

Inference verification is the task of verifying that a piece of text was actually generated by a specific model with specific sampling parameters. We use Token-DiFR to accomplish inference verification.

Why does this matter?

Inference verification has two key applications:

  • Detecting steganography — identifying covert channels hidden in LLM outputs (arXiv:2511.02620)
  • Verifying inference providers — ensuring a provider actually ran the claimed model and didn't substitute a cheaper one (arXiv:2511.20621)

Importantly, for inference verification to be trustworthy, you also need to trust the verifier itself. That's why we run this inside a Trusted Execution Environment (TEE) — the verification code runs in hardware-isolated memory that even the server operator cannot inspect or tamper with.

Verify Local Text

This example text is stored on the device. Click "Run Verification" to verify it was generated by the model.

Prompt
Response Text to Verify

1. Query OpenRouter

Simulate a dishonest provider

Consider: you requested a response from Llama 3.1 8B, but the provider wants to save money so they serve you Llama 3.2 3B instead. Click below to simulate such a query — verification should flag the mismatch.

Results