---
sidebar_position: 5.6
tags: [security]
---

> ## Documentation Index
> Fetch the complete documentation index at: https://docs.lium.io/llms.txt
> Use this file to discover all available pages before exploring further.
>
> Lium (lium.io) is a GPU rental marketplace: rent NVIDIA GPUs by the hour, billed per second.

# Private models and datasets

Run a model or dataset the GPU provider must not see. You rent a Confidential Computing machine, the pod proves what machine it is on, and only then does your own key service release a short-lived key that decrypts your data.

## What is guaranteed

On a **Confidential Computing** machine, after your verifier checks the pod's quote ([step 4](#4-verify-the-quote-and-release-the-key)):

- **The provider cannot read the VM's memory or disk.** Intel TDX encrypts them.
- **The quote does not identify your image or your pod.** It measures the machine. Rent every GPU so yours is the only pod on it.
- **Lium's node software still runs inside the VM.** The VM keeps the provider out, not Lium.
- **GPU memory and the CPU-GPU link are not covered.** The label comes from the CPU's TDX quote; Lium only observes GPU attestation. If you need GPU protection, check the GPUs with NVIDIA's attestation tools from your pod before you load anything.
- **Data that leaves the VM** (backups, volumes, logs, network) is protected only if you encrypt it.

On an ordinary machine none of this holds ([Pod security](./security)).

## The flow

**1. Prepare outside Lium**

- Encrypt the weights or dataset with a data key and upload only the ciphertext.
- Keep the data key in a key service you control (the **key broker**).
- Build an image with your code and the [dstack SDK](https://github.com/Dstack-TEE/dstack), and no secrets.

**2. Rent a Confidential Computing machine**

On **Browse Pods**, turn on the **Confidential Computing** filter and rent **all** GPUs of a machine. With the SDK: `Lium.up(executor_id=..., gpu_count=...)`; with the CLI: `lium up <id> -c <gpu_count>`.

Don't put secrets in the rent request, template or start command.

**3. Get a quote from inside the pod**

Make an in-memory key pair, get a nonce from your broker, and bind both into the quote:

```python
import hashlib

from cryptography.hazmat.primitives.asymmetric.x25519 import X25519PrivateKey
from cryptography.hazmat.primitives.serialization import Encoding, PublicFormat
from dstack_sdk import DstackClient

nonce = get_nonce_from_your_broker()  # your code
private_key = X25519PrivateKey.generate()
public_key = private_key.public_key().public_bytes(Encoding.Raw, PublicFormat.Raw)

quote = DstackClient().get_quote(hashlib.sha256(public_key + nonce).digest())
send_to_your_broker(quote.quote, quote.event_log, public_key, nonce)  # your code
```

### 4. Verify the quote and release the key {#4-verify-the-quote-and-release-the-key}

The broker releases nothing until all three pass:

1. **Signature and TCB:** the quote verifies with an Intel DCAP verifier or the dstack verifier, and its TCB status is up to date. Refuse a quote whose TCB is out of date.
2. **Binding:** `report_data` equals `sha256(public_key + nonce)`, with a fresh, unused nonce.
3. **Measurements:** MRTD and RTMR0–2 match the dstack OS release in the [CVM guide](../providers/nodes/cvm.md), and RTMR3 matches the "CVM attestation" section of the [executor release notes](https://github.com/Datura-ai/lium-io/releases).

Then the broker encrypts the data key to the pod's public key and returns it.

**5. Load the data**

Decrypt into memory or `/root`. Don't write plaintext to a volume or a backed-up path.

**6. End the run**

Terminate the pod. Revoke any storage credential still valid.

## Keep credentials short-lived

- **Lium API key:** on your laptop or CI, never the pod. Set a [budget](./api-keys#budgets).
- **Storage access:** a read-only pre-signed URL or temporary token that expires on its own.
