README.md

Model Signing

This project demonstrates how to protect the integrity of a model by signing it with Sigstore, a tool for making code signatures transparent without requiring management of cryptographic key material.

When users download a given version of a signed model they can check that the signature comes from a known or trusted identity and thus that the model hasn't been tampered with after training.

Signing events are recorded to Sigstore's append-only transparency log. Transparency logs make signing events discoverable: Model verifiers can validate that the models they are looking at exist in the transparency log by checking a proof of inclusion (which is handled by the model signing library). Furthermore, model signers that monitor the log can check for any unexpected signing events.

Model signers should monitor for occurences of their signing identity in the log. Sigstore is actively developing a log monitor that runs on GitHub Actions.

Usage

You will need to install a few prerequisites to be able to run all of the examples below:

sudo apt install git git-lfs python3-venv python3-pip unzip
git lfs install

After this, you can clone the repository, create a Python virtual environment and install the dependencies needed by the project:

git clone git@github.com:sigstore/model-transparency.git
cd model-transparency/model_signing
python3 -m venv test_env
source test_env/bin/activate
os=Linux # Supported: Linux, Windows, Darwin.
python3 -m pip install --require-hashes -r "install/requirements_${os}".txt

After this point, you can use the project to sign and verify models and checkpoints. A help message with all arguments can be obtained by passing -h argument, either to the main driver or to the two subcommands:

python3 main.py -h
python3 main.py sign -h
python3 main.py verify -h

Signing a model requires passing an argument for the path to the model. This can be a path to a file or a directory (for large models, or model formats such as SavedModel which are stored as a directory of related files):

path=path/to/model
python3 main.py sign --path "${path}"

The sign process will start an OIDC workflow to generate a short lived certificate based on an identity provider. This will be relevant when verifying the signature, as shown below.

Note: The signature is stored as <file>.sig for a model serialized as a single file, and <dir>/model.sig for a model in a folder-based format.

For verification, we need to pass both the path to the model and identity related arguments:

python3 main.py verify --path "${path}" \
    --identity-provider https://accounts.google.com \
    --identity myemail@gmail.com

For developers signing models, there are three identity providers that can be used at the moment:

• Google's provider is https://accounts.google.com.
• GitHub's provider is https://github.com/login/oauth.
• Microsoft's provider is https://login.microsoftonline.com.

For automated signing using a workload identity, the following platforms are currently supported, shown with their expected identities:

• GitHub Actions (https://github.com/octo-org/octo-automation/.github/workflows/oidc.yml@refs/heads/main)
• GitLab CI (https://gitlab.com/my-group/my-project//path/to/.gitlab-ci.yml@refs/heads/main)
• Google Cloud Platform (SERVICE_ACCOUNT_NAME@PROJECT_ID.iam.gserviceaccount.com)
• Buildkite CI (https://buildkite.com/ORGANIZATION_SLUG/PIPELINE_SLUG)

Supported Models

The library supports multiple models, from multiple training frameworks and model hubs.

For example, to sign and verify a Bertseq2seq model, trained with TensorFlow, stored in TFHub, run the following commands:

model_path=bertseq2seq
wget "https://tfhub.dev/google/bertseq2seq/bert24_en_de/1?tf-hub-format=compressed" -O "${model_path}".tgz
mkdir -p "${model_path}"
cd "${model_path}" && tar xvzf ../"${model_path}".tgz && rm ../"${model_path}".tgz && cd -
python3 main.py sign --path "${model_path}"
python3 main.py verify --path "${model_path}" \
    --identity-provider https://accounts.google.com \
    --identity myemail@gmail.com

For models stored in Hugging Face we need the large file support from git, which can be obtained via

sudo apt install git-lfs
git lfs install

After this, we can sign and verify a Bert base model:

model_name=bert-base-uncased
model_path="${model_name}"
git clone --depth=1 "https://huggingface.co/${model_name}" && rm -rf "${model_name}"/.git
python3 main.py sign --path "${model_path}"
python3 main.py verify --path "${model_path}" \
    --identity-provider https://accounts.google.com \
    --identity myemail@gmail.com

Similarly, we can sign and verify a Falcon model:

model_name=tiiuae/falcon-7b
model_path=$(echo "${model_name}" | cut -d/ -f2)
git clone --depth=1 "https://huggingface.co/${model_name}" && rm -rf "${model_name}"/.git
python3 main.py sign --path "${model_path}"
python3 main.py verify --path "${model_path}" \
    --identity-provider https://accounts.google.com \
    --identity myemail@gmail.com

We can also support models from the PyTorch Hub:

model_name=hustvl/YOLOP
model_path=$(echo "${model_name}" | cut -d/ -f2)
wget "https://github.com/${model_name}/archive/main.zip" -O "${model_path}".zip
mkdir -p "${model_path}"
cd "${model_path}" && unzip ../"${model_path}".zip && rm ../"${model_path}".zip && shopt -s dotglob && mv YOLOP-main/* . && shopt -u dotglob && rmdir YOLOP-main/ && cd -
python3 main.py sign --path "${model_path}"
python3 main.py verify --path "${model_path}" \
    --identity-provider https://accounts.google.com \
    --identity myemail@gmail.com

We also support ONNX models, for example Roberta:

model_name=roberta-base-11
model_path="${model_name}.onnx"
wget "https://github.com/onnx/models/raw/main/text/machine_comprehension/roberta/model/${model_name}.onnx"
python3 main.py sign --path "${model_path}"
python3 main.py verify --path "${model_path}" \
    --identity-provider https://accounts.google.com \
    --identity myemail@gmail.com

Benchmarking

Install as per Usage section. Ensure you have enough disk space:

• if passing 3rd script argument as true: at least 50GB
• otherwise: at least 100GB

To run the benchmarks:

git clone git@github.com:sigstore/model-transparency.git
cd model-transparency/model_signing
bash benchmarks/run.sh https://accounts.google.com myemail@gmail.com [true]

A single run was performed.

Hashes used:

• H1: Hashing using a tree representation of the directory.
• H2: Hashing using a list representation of the directory. (Implementation is parallized with shards of 1GB sizes across vCPUs).

Machine M1: Debian 6.3.11 x86_64 GNU/Linux, 200GB RAM, 48 vCPUs, 512KB cache, AMD EPYC 7B12:

Hash	Model	Size	Sign Time	Verify Time
H1	roberta-base-11	8K	0.8s	0.6s
H1	hustvl/YOLOP	215M	1.2s	0.8s
H1	bertseq2seq	2.8G	4.6s	4.4s
H1	bert-base-uncased	3.3G	5s	4.7s
H1	tiiuae/falcon-7b	14GB	12.2s	11.8s
H2	roberta-base-11	8K	1s	0.6s
H2	hustvl/YOLOP	215M	1s	1s
H2	bertseq2seq	2.8G	1.9s	1.4s
H2	bert-base-uncased	3.3G	1.6s	1.1s
H2	tiiuae/falcon-7b	14GB	2.1s	1.8s

Machine M2: Debian 5.10.1 x86_64 GNU/Linux, 4GB RAM, 2 vCPUs, 56320 KB, Intel(R) Xeon(R) CPU @ 2.20GHz:

Hash	Model	Size	Sign Time	Verify Time
H1	roberta-base-11	8K	1.1s	0.7s
H1	hustvl/YOLOP	215M	1.9s	1.7s
H1	bertseq2seq	2.8G	18s	23.2s
H1	bert-base-uncased	3.3G	23.4s	18.9s
H1	tiiuae/falcon-7b	14GB	2m4s	2m2s
H2	roberta-base-11	8K	1.1s	0.8s
H2	hustvl/YOLOP	215M	1.9s	1.6s
H2	bertseq2seq	2.8G	13.8s	25.9s
H2	bert-base-uncased	3.3G	22.7s	23.3s
H2	tiiuae/falcon-7b	14GB	2m.1s	2m3s