Getting started with AICert!

To get started, we will walk you through the steps needed to configure and launch AICert using an example where we finetune a TinyLlama-1.1B model with proof that the model was derived from the official TinyLlama-1.1B pre-trained model, our open-source training code and the dataset that we specify.

The end-user can then use our Python client SDK to verify that the AICert proof file is genuine and check the inputs used (pre-trained model, dataset, and code) to produce our model.

All the files shown in this example are available on Github.

The workflow is as follows:

workflow

AI builder workflow 🛠️

The AI builder prepares an axolotl configuration file containing the name of the pretrained model, dataset and finetuning parameters
The AI builder launches AICert using the CLI tool

AICert workflow ⚙️

AICert provisions a VM with the required hardware/software stack
AICert executes the finetuning as described in the configuration file
AICert returns a link to access the training outputs and a cryptographic proof file

The proof file contains measurements relating to the software stack, the training code and inputs and the training outputs (e.g. the trained model)

End user workflow 👩🏻‍💻

The end user verifies the certificate and the inputs and configuration used to create the trained model

End users will only be able to verify the inputs and configuration where they have access to the original data

Let’s now take a look at the steps that the AI builder and end users must follow in more detail.

AI builder POV: creating an AI model certificate

Step 1: Preparing the OS image (Only performed once)

Step 1-A: Set Azure configuration

We must first configure the Azure region and resource names before we begin creating the Mithril OS image.

AZ_RESOURCE_GROUP="your_resource_group"
AZ_REGION="your_region"
GALLERY_NAME="your_gallery_name"
STORAGE_ACCOUNT="your_storage"
STORAGE_CONTAINER="your_container"

The size of the Azure VM can be set in the file: variables.tf

variable "instance_type" {
  type        = string
  default     = "Standard_NC24ads_A100_v4"
  description = "Type/Size of VM to create."
}

The default size is Standard_NC24ads_A100_v4

Step 1-B: Create the Mithril OS image

We use mkosi to create an OS image containing the required containers for finetuning and attestation. The OS is a minimal image with all unnecessary services and modules removed and network restrictions in place.

We've grouped a few steps into a script named create_MithrilOS.sh. This script creates the Mithril OS in a disk, uploads it to Azure and converts it into an image, and finally generates the measurements of the OS (which attest the contents of the disk)

# log in to Azure CLI
az login

# create Mithril OS image
./create_MithrilOS.sh

Mithril OS contains the following containers:

aicert-server: This container contains the aicert server, it is the server the client communicates with.
aicert-base: This is a minimal image that is used to download the model and dataset specified in the configuration.
caddy: Caddy is a reverse-proxy container.
axolotl: Axolotl is a container containing the axolotl finetuning framework.

Inputs (pre-trained model, datasets) are downloaded from huggingface as specified in the AICert config file. We consider these as project resources and individually hash these files.

The whole repo will be moved to a /workspace/src folder within our Docker container.

Step 2: Modifying the AICert config file

We have a default config in the axolotl_yaml folder in the client. You can either modify this file or create your own configuration yaml file.

Step 3: Launching the traceable training

Finally, to launch the traceable training process and get back our AI certificate, we can use the AICert CLI tool and run the aicert finetune command.

aicert finetune

AICert will look for a yaml file named aicert.yaml in the current working directory and load it. If you want to use a configuration file with another name, you may use the config option (it must be in the current working directory).

aicert finetune --config custom_config.yaml

Once the training process is complete, we obtain an attestation report named attestation.json, which binds the hashes of the weights with the training code and inputs, as well as the software stack of the VM used for training.

This proof file can now be shared with outside parties to prove to them the model comes from using the specified training code and data.

End user POV: Verifying and inspecting the AI certificate

The end user can then use our Python SDK to:

Verify that the AICert proof file is legitimate
Verify the software stack of the VM used for training
Verify the inputs used for training against known hash values

Verification of the AI certificate and VM software stack

End users can verify the exported proof file is genuine and does contain any unexpected measurements by using the verify() method provided by the AICert Python package with no arguments.

aicert verify

The verify() method checks two things:

The authenticity of the certificate's signature
The validity of the hashed values of the whole software stack or boot chain of the VM used to train the dataset. This guarantees that the certification process is valid and not compromised.

Warning

The verify() method does not attest that the base model and dataset are trustworthy. They have to be audited independently. However, if the certification process is valid, the AI builder can now be held accountable- if they use, for instance, poisoned data to train the model, this can be verified a posteriori.

If the proof file contains a false signature or any false values, an error will be raised. False hashed values could signal that the software stack of the VM used for training was misconfigured or even tampered with.

If the verify() method does not return any errors, it means that the AI certificate is genuine.

If the client is run on a different machine than the one used to generate the OS disk, the following measurement files must be shared:

container_measurements.json
measurements_azure.json
measurements_qemu.json (only required if the OS is being run locally for testing)

Place these files in the security_config folder in the client.

Inspecting the proof file

We can use the proof file to manually check the hashed values of the model's inputs or output hash against known values.

For example, an attestation report loks like this:

{
    "ca_cert": "",
    "event_log": [
        "{\"event_type\": \"axolotl_configuration\", \"content\": {\"spec\": {\"config_file\": {\"adapter\": \"lora\", \"base_model\": \"model/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\", \"bf16\": \"auto\", \"dataset_prepared_path\": null, \"datasets\": [{\"path\": \"dataset/mhenrichsen/alpaca_2k_test\", \"type\": \"alpaca\"}], \"debug\": null, \"deepspeed\": null, \"early_stopping_patience\": null, \"eval_sample_packing\": false, \"evals_per_epoch\": 4, \"flash_attention\": false, \"fp16\": null, \"fsdp\": null, \"fsdp_config\": null, \"gradient_accumulation_steps\": 4, \"gradient_checkpointing\": true, \"group_by_length\": false, \"learning_rate\": 0.0002, \"load_in_4bit\": false, \"load_in_8bit\": true, \"local_rank\": null, \"logging_steps\": 1, \"lora_alpha\": 16, \"lora_dropout\": 0.05, \"lora_fan_in_fan_out\": null, \"lora_model_dir\": null, \"lora_r\": 32, \"lora_target_linear\": true, \"lr_scheduler\": \"cosine\", \"micro_batch_size\": 2, \"model_type\": \"LlamaForCausalLM\", \"num_epochs\": 4, \"optimizer\": \"adamw_bnb_8bit\", \"output_dir\": \"./lora-out\", \"pad_to_sequence_len\": true, \"resume_from_checkpoint\": null, \"sample_packing\": true, \"saves_per_epoch\": 1, \"sequence_len\": 4096, \"special_tokens\": null, \"strict\": false, \"tf32\": false, \"tokenizer_type\": \"LlamaTokenizer\", \"train_on_inputs\": false, \"val_set_size\": 0.05, \"wandb_entity\": null, \"wandb_log_model\": null, \"wandb_name\": null, \"wandb_project\": null, \"wandb_watch\": null, \"warmup_steps\": 10, \"weight_decay\": 0.0, \"xformers_attention\": null}}, \"resolved\": {\"hash\": \"156cd3d4b6f38dda838dbf7e95f8935268f91f091b1af3209f3a4239eea3db9f\"}}}",
        "{\"event_type\": \"input_image\", \"content\": {\"spec\": {\"image_name\": \"@local/aicert-base:latest\"}, \"resolved\": {\"id\": \"sha256:2f1361151ba80d5d0493463b63330c2a6f00e8d92e449c7d6a6febb9a03dafc1\"}}}",
        "{\"event_type\": \"input_resource\", \"content\": {\"spec\": {\"resource_proto\": {\"resource_type\": \"model\", \"repo\": \"https://huggingface.co/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\", \"hash\": \"036fa4651240b9a1487f709833b9e4b96b4c1574\", \"path\": \"model/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\"}}, \"resolved\": {\"hash\": \"sha1:b'036fa4651240b9a1487f709833b9e4b96b4c1574\\\\n'\"}}}",
        "{\"event_type\": \"input_resource\", \"content\": {\"spec\": {\"resource_proto\": {\"resource_type\": \"dataset\", \"repo\": \"https://huggingface.co/datasets/mhenrichsen/alpaca_2k_test\", \"hash\": \"d05c1cb585e462b16532a44314aa4859cb7450c6\", \"path\": \"dataset/mhenrichsen/alpaca_2k_test\"}}, \"resolved\": {\"hash\": \"sha1:b'd05c1cb585e462b16532a44314aa4859cb7450c6\\\\n'\"}}}",
        "{\"event_type\": \"input_image\", \"content\": {\"spec\": {\"image_name\": \"@local/axolotl:latest\"}, \"resolved\": {\"id\": \"sha256:dcc5b2108616a303e544faf68a060304fe514788e72e17264966bebb5ef99062\"}}}",
        "{\"event_type\": \"outputs\", \"content\": [{\"spec\": {\"path\": \"finetuned-model.zip\"}, \"resolved\": {\"hash\": \"8739c76e681f900923b900c9df0ef75cf421d39cabb54650c4b9ad19b6a76d85\"}}]}"
    ],
    "remote_attestation": {
        "quote": {
            "message": {
                "base64": "/1RDR4AYACIAC14LdKX3smznOCoo8tCef86t1BXz8V9Vv54XgYYGs4cSAAAAAAAAABGlJgAAAAIAAAAAASAgAxIAEgADAAAAAQALA////wAgz8Uc1hRbl+IPWqW0XVjdXr6CsNgjo0V877qEZ2mcAcc="
            },
            "signature": {
                "base64": "ABQACwEAGretfAiEAr2t1yHYHI0ACELTgttOANErHrmPaUf26QH6kWElEBwaW4JsqPusiGl3q4eh8QIe3J5qKm8mHTvbhlSIgIfav+K2xrjFvAjDi/qeuwWmO5oYhDSCbnNlFZ1BQU3396Wj6k8ebAL8dsHbWFqyJFcZ4LJGnmU/YIF3rAw4g9689LNn2ogK/a6pdfyH/IuVxWfM0Mn4C7GnchHc6gspDNGjnF4f486dAoErIsaULfDhbvKzmoHHsW9amrOGBBHSWH6t491q4NEGeYwtWc5kpeuJreVA2JHP/7z1dTGExg7zLXCBcrwDSzDmWyav7gEzDGFcXY6pRhojqQxF/g=="
            },
            "pcr": {
                "base64": "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"
            }
        },
        "cert_chain": [
            {
                "base64": "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"
            },
            {
                "base64": "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"
            },
            {
                "base64": "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"
            }
        ]
    }
}