Перейти к содержанию

🔄 Авто-синхронизация: из Discussion #951 каждые 6 часов.

TEE Implementation

Автор: @mtvnastya · Категория: 💡 Proposals · Создано: 2026-03-26 05:10 UTC · Обновлено: 2026-06-05 15:05 UTC

📝 Описание

TEE

This proposal outlines the integration of Trusted Execution Environments (TEE) into the Gonka network architecture. By leveraging hardware-level isolation (Intel TDX, AMD SEV-SNP, NVIDIA Confidential Compute), we can protect user payloads and inference data from the physical hosts/miners, enabling true "Confidential MLNodes." This document proposes changes to the node registry, validation pipeline, and economic incentives.

Problem

Currently, user requests data sent to the Gonka network is hidden from the public but transparent to the specific host (miner) executing and validating the workload. Even with off-chain payloads, a malicious operator with physical access to the machine can access the data by using custom binaries or modifying the execution environment.

For enterprise and privacy-sensitive adoption, trust in the protocol must supersede trust in the hardware operator.

Proposal

We propose a new node class: the Confidential MLNode. This node operates within a TEE, isolating the full Virtual Machine (VM) from the hypervisor.

Supported CPUs:

  • Intel TDX: https://www.intel.com/content/www/us/en/developer/tools/trust-domain-extensions/overview.html
  • AMD SEV-SNP: https://www.amd.com/en/developer/sev.html

Both technologies support attaching NVIDIA GPUs with Confidential Compute: - https://docs.nvidia.com/nvidia-secure-ai-with-blackwell-and-hopper-gpus-whitepaper.pdf

A VM with an attached GPU can encrypt all in-memory data, data transferred to the GPU, and data in GPU memory from the hypervisor.

By carefully modifying how MLNode works with data: - Inferences are decrypted only in-memory - No sensitive data is written to disk - No open TCP ports are used for unencrypted data transfer - No sensitive data appears in logs

The full inference pipeline can be protected from both the host/miner and the hypervisor (server owner in the case of rented servers).

Using a temporary public key generated by the TEE-protected MLNode enables end-to-end data encryption without any decrypted data existing in unprotected memory.

Architecture Upgrade

MLNode

  • VM with MLNode is launched from an image with metadata saved on-chain (model embedded). Full MLNode behavior is defined by its REST API
  • VM generates a key pair for data encryption/decryption

  • VM provides an attestation certificate signed by the CPU's hardware module. The signed data includes:

    • Hardware information
    • GPU information, including Confidential Compute enabled, exclusive access to GPU, and GPU attestation
    • Exact VM metadata
    • Public key from generated key pair
    • Proof that VM launched from correct entrypoint (target application)

    Note: If the VM is restarted or the image is replaced, the private key will be lost. - Host publishes this certificate on-chain

Open Question 1: Should the certificate be validated on-chain?

All requests to the Confidential MLNode are encrypted using the public key from the certificate.

Relevant materials:

Framework to simplify TEE deployment: - https://github.com/Dstack-TEE/dstack

Insightful paper about NVIDIA Confidential Compute: - https://arxiv.org/pdf/2507.02770

Network Node

1. New MLNode Type

Confidential MLNode with a separate pipeline for scheduling inferences. These nodes do not require inference validation.

Confidential MLNodes have an associated attestation certificate and public key.

2. Certificate Validation

Requirements for certificate validation: - Publishing certificates to the blockchain - Maintaining Intel and AMD root certificates (Root of Trust keys) for verification - Maintaining a list of secure software versions and excluded (untrusted) certificates

Open Question 2: Should full certificate validation be performed on-chain? Probably yes, on recording.

3. Inference Pipeline

Current /chat/completions request flow:

Dev -> Transfer Agent (Network Node) -> Executor (Network Node) -> Executor (MLNode)
image (64)

https://what-is-gonka.hashnode.dev/decentralized-ai-inference-balancing-security-and-performance

In the current system, any request is scheduled to any executor. For Confidential MLNodes, the workflow changes:

  • User obtains the Confidential MLNode's public key for encryption
  • User encrypts the payload using the public key
  • Encrypted request is sent to and decrypted on the specific Confidential MLNode

This design enables users to send /chat/completions requests directly to the Confidential MLNode without proxying through multiple Network Nodes.

Possible Pipeline & Payment

Proposed workflow:

  1. User sends a request to any Network Node to reserve a Confidential MLNode with quota (possibly for a long-living session)
  2. Chain randomly assigns a Confidential MLNode, moves quota from user's account to escrow, returns the Confidential MLNode's public key and certificate to the user
  3. User encrypts payload and sends to MLNode:
  4. Option 1: Directly to MLNode
  5. Option 2: Through any Network Node in the network
  6. Confidential MLNode performs computation and encrypts response with the user's public key
  7. User decrypts response with their private key
  8. Confidential MLNode signs response metadata (used tokens) and sends to its Network Node
  9. Network Node submits signed metadata to claim coins from escrow. Confidential MLNode signature is verified on-chain before claim

Note: Metadata can be sent in batches periodically to optimize on-chain transactions. Unclaimed quota is automatically refunded at the next epoch boundary.

Signed metadata from a TEE key is inherently trusted - the execution environment is pre-defined and verified via attestation. The MLNode cannot produce valid signatures without running in the attested TEE, which guarantees the correct binary ran.

options (1)

Open Question 3: How to provide redundancy when a Confidential MLNode becomes unavailable?

Economic Incentive

  • Confidential MLNode inferences do not need validation => eliminating the need to share work rewards with validators => eliminating need to generate and store artifacts for validation
  • Confidential MLNodes do not need to participate in all PoC rounds if they maintain the same key pair, since this proves consistent hardware
  • Confidential Inferences are priced in a separate dynamic pricing group per model, making these inferences more profitable when demand is high

Open Question 4: Since Confidential MLNodes have less validation overhead, should they receive higher bitcoin-style rewards to incentivize enabling TEE?

💬 Комментарии (2)

Комментарий 1 — @x0152

2026-05-25 21:21 UTC

Some first experiments on this in #1246. Not a full implementation, but first end-to-end wiring across all layers:

  • Chain: on-chain attestation flow (submit + peer cross-validation, verification runs in dapi)
  • DAPI: intel-tdx-lite verifier (offline), HPKE envelope client, attestation workers, encrypted forwarding to ml-node
  • ml-node: HPKE-encrypted inference endpoint, dstack derived keys, TDX quote with MRTD
  • devshard: encrypted transport for ml-node forwarding
  • Phala CVM smoke test notebook

Could be a useful starting point. Happy to answer questions

Комментарий 2 — @Otli4nikDe

2026-06-05 15:05 UTC

What if we rethink the problem for the entire Gonka network and reframe it like this: We accept as a given that the data is exposed on the hosts (!). This means anyone can potentially see it. But how can we ensure that when viewing this data, two key conditions are met:

  1. No real leaks of private data — even if API keys or other credentials are present.
  2. When viewing the context, there is maximum anonymity / fakeness?!

Then the question becomes: what’s the point if I read someone’s email, but that email address doesn’t actually exist? Or if I get hold of keys, but they are fake?!

## Alternative tier: a lightweight "Surrogate-Scrubbing Gateway" in a CPU TEE (no confidential GPU required)

First — big +1 on the Confidential MLNode direction. Full TEE inference (TDX/SEV-SNP + NVIDIA CC) is the strongest guarantee and the right long-term target. I'd like to propose a complementary, lower-cost confidentiality tier that could ship sooner and run on the existing GPU fleet.

### The idea

Instead of running the whole model inside a confidential GPU, put a tiny scrub/restore proxy inside a CPU-only confidential VM (TDX or SEV-SNP) and leave inference on ordinary, non-confidential MLNodes:

image

  • The client sends data into the Gonka network/chain, where sensitive information is substituted (scrubbed) at some point inside the chain — either at the entry point or before it reaches the LLM Node. (same attestation/on-chain certificate model as the main proposal — we just attest the gateway, not the inference node).
  • Inside the TEE, the gateway detects sensitive entities (API keys, credit cards, SSNs, IBANs, crypto addresses, emails, etc.) and replaces them with format-preserving surrogates — realistic fake values of the same shape, not ciphertext, so the model can still reason over plausible structure.
  • The surrogated prompt goes to a normal MLNode for inference at full speed.
  • On the way back, the gateway restores surrogates → originals inside the TEE, before returning the (re-encrypted) response to the client. Bidirectional.
  • The surrogate↔original vault lives only inside the TEE, in memory, and is ephemeral (lost on VM restart) — consistent with the ephemeral-key model already proposed here.

What’s the implementation cost?

First of all, I look at it from the user’s perspective — and that means latency! Where and how much time will be spent when using the encryption/scrubbing gateway?I’ll share real test results below. I use exactly the same approach for my own Personal Assistants, except my code is not in Go, but compiled as a Rust binary.Categories I filter for:

  • API_KEY
  • Credit Cards
  • US SSN
  • Crypto (bc1/1/3, 0x***)
  • IP (IPv4 + loose IPv6)
  • Email
  • Passwords

Test conditions: Context = 800 KB ≈ 200k tokens Results: 1. With 200k tokens and 120 entities→ scan + replace time = 2.36 ms 2. With 200k tokens and 2022 entities → scan + replace time = 5.41 ms *200k tokens is close to the maximum context for top models like Kimi & MiniMax 2.7.

Conclusion: The added latency for processing one prompt in both directions can be considered negligible compared to the actual inference time on the GPUs. Much more important is how to architect this. I see two main approaches: 1) Add another node between the Network Node and the LLM Node (GPU). → This will add network latency (roughly up to 300 ms). 2) Run TEE by default on the Network Nodes themselves (on CPU). → In this case there is almost zero added latency, and the hosts will always receive only surrogate/fake data :)

💬 Комментарии