# Model Transparency
<!-- markdown-toc --bullets="-" -i README.md -->
<!-- toc -->
- [Overview](#overview)
- [Model Signing](#model-signing)
- [Model Signing CLI](#model-signing-cli)
- [Model Signing API](#model-signing-api)
- [Model Signing Format](#model-signing-format)
- [Status](#status)
- [Contributing](#contributing)
<!-- tocstop -->
## Overview
There is currently significant growth in the number of ML-powered applications.
This brings benefits, but it also provides grounds for attackers to exploit
unsuspecting ML users.
Building on the work with [Open Source Security Foundation][openssf], we are
creating this collection of projects to strengthen the ML supply chain in
_the same way_ as the traditional software supply chain.
The focus is on providing *verifiable* claims about the integrity and provenance
of the resulting models, meaning users can check for themselves that these
claims are true rather than having to just trust the model trainer.
## Model Signing
This project demonstrates how to protect the integrity of a model by signing it.
We support generating signatures via [Sigstore](https://www.sigstore.dev/), a
tool for making code signatures transparent without requiring management of
cryptographic key material. But we also support traditional signing methods, so
models can be signed with public keys or signing certificates as well as
PKCS #11 enabled devices *(install with `pip install model-signing[pkcs11]` to enable this functionality)*.
The signing part creates a
[sigstore bundle](https://github.com/sigstore/protobuf-specs/blob/main/protos/sigstore_bundle.proto)
protobuf that is stored as in JSON format. The bundle contains the verification
material necessary to check the payload and a payload as a
[DSSE envelope](https://github.com/sigstore/protobuf-specs/blob/main/protos/envelope.proto).
Further the DSSE envelope contains an in-toto statment and the signature over
that statement. The signature format and how the the signature is computed can
be seen
[here](https://github.com/secure-systems-lab/dsse/blob/v1.0.0/protocol.md).
Finally, the statement itself contains subjects which are a list of (file path,
digest) pairs a predicate type set to `https://model_signing/signature/v1.0` and
a dictionary of predicates. The idea is to use the predicates to store (and
therefor sign) model card information in the future.
The verification part reads the sigstore bundle file and firstly verifies that the
signature is valid and secondly compute the model's file hashes again to compare
against the signed ones.
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.
When using Sigstore, 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](https://github.com/sigstore/rekor-monitor) that runs on GitHub Actions.
### Model Signing CLI
After installing the package, the CLI can be used via either `python -m
model_signing <args>` or by calling the binary directly, `model_signing <args>`.
Users that don't want to install the package, but want to test this using the
repository can do the same using [Hatch](https://hatch.pypa.io/latest/) via
`hatch run python -m model_signing <args>`.
For the remainder of the section, we would use `model_signing <args>` method.
The CLI has two subcommands: `sign` for signing and `verify` for verification.
Each subcommand has another level of subcommands to select the signing method
(`sigstore` -- the default, can be skipped --, `key`, `certificate`). Then, each
of these subcommands has several flags to configure parameters for
signing/verification.
For the demo, we will use the `bert-base-uncased` model, which can be obtained
via:
```bash
[...]$ git clone --depth=1 "https://huggingface.co/bert-base-uncased"
```
We remove the `.git` directory since that should not be included in the
signature:
```bash
[...]$ rm -rf bert-base-uncased/.git
```
By default, the code also ignores git related paths.
The simplest example of the CLI is to sign a model using Sigstore:
```bash
[...]$ model_signing sign bert-base-uncased
```
This will open an OIDC flow to obtain a short lived token for the certificate.
The identity used during signing and the provider must be reused during
verification.
To only compute and output the digest of the model, you can use the `digest`
subcommand, pointing it to the model directory:
```bash
[...]$ model_signing digest bert-base-uncased
```
The digest subcommand follows the same ignore rules used when signing.
## Using Private Sigstore Instances
To use a private Sigstore setup (e.g. custom Rekor/Fulcio), use the `--trust-config` flag:
```bash
[...]$ model_signing sign bert-base-uncased --trust-config client_trust_config.json
```
For verification:
```bash
[...]$ model_signing verify bert-base-uncased \
--signature model.sig \
--trust-config client_trust_config.json
--identity "$identity"
--identity-provider "$oidc_provider"
```
The `client_trust_config.json` file should include:
- A signed target trust root
- A `signingConfig` section with your private Rekor, Fulcio, and CT log endpoints
- Public keys for verification (if applicable)
You can find an example `client_trust_config.json` that references the public Sigstore production services in the Sigstore Python repository [here](https://github.com/sigstore/sigstore-python/blob/main/test/assets/trust_config/config.v1.json).
As another example, here is how we can sign with private keys. First, we
generate the key pair:
```bash
[...]$ openssl ecparam -name prime256v1 -genkey -noout -out key.priv
[...]$ openssl ec -in key.priv -pubout > key.pub
```
And then we use the private key to sign.
```bash
[...]$ model_signing sign key bert-base-uncased --private-key key.priv
```
All signing methods support changing the signature name and location via the
`--signature` flag:
```bash
[...]$ model_signing sign bert-base-uncased --signature model.sig
```
Consult the help for a list of all flags (`model_signing --help`, or directly
`model_signing` with no arguments)
On verification we use the `verify` subcommand. To verify a Sigstore signed
model we use
```bash
[...]$ model_signing verify bert-base-uncased \
--signature model.sig \
--identity "$identity" \
--identity-provider "$oidc_provider"
```
Where `$identity` and `$oidc_provider` are those set up during the signing flow
and `--signature` must point to the signature to verify.
For developers signing models with Sigstore, 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`.
* GitHub Actions uses `https://token.actions.githubusercontent.com`
* 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`)
Similarly, for key verification, we can use
```bash
[...]$ model_signing verify key bert-base-uncased \
--signature resnet.sig --public-key key.pub
```
#### Signing with PKCS #11 URIs
Signing with PKCS #11 enabled crypto devices is supported through RFC 7512
compliant PKCS #11 URIs. The URI can be used in place of the private key
when siging with a private key or certificate.
The following features are supported/required:
- The PKCS #11 URI must either provide the module through the query
parameter 'module-path', or the query parameter 'module-name' must
describe the name of a module that can be found in well-known
directories of Linux distributions.
- A token can be selected based on a provided 'slot-id' path parameter.
The first token that matches the given slot-id will be used. If a
'token' path parameter is also provided, then it will be used for
selecting the appropriate token by its label.
- When no 'slot-id' is given then all slots are searched for by the
name of the given 'token'.
- A PIN may be provided as 'pin-value' query parameter or may be read
from a file described by the 'pin-source' query parameter.
- An 'id' path parameter and/or key label (path parameter 'object') must
be provided to select the signing key.
- The public key on the PKCS #11 device will also be accessed during
signing.
- The signing key must be of type NIST P256/384/521 (secp256/384/512r1).
With a PKCS #11 URI describing the private key, we can use the following
for signing:
```bash
[...]$ model_signing sign pkcs11-key --signature model.sig \
--pkcs11_uri "pkcs11:..." /path/to/your/model
```
For signature verification it is necessary to retrieve the public key from
the PKCS #11 device and store it in a file in PEM format. With can then use:
```bash
[...]$ model_signing verify key --signature model.sig\
--public-key key.pub /path/to/your/model
```
#### OpenTelemetry Support
Model signing supports optional distributed tracing and observability through OpenTelemetry. This allows you to monitor signing operations, track performance, and integrate with observability platforms.
To enable OpenTelemetry support, install the optional dependencies:
```bash
pip install model-signing[otel]
```
Once installed, OpenTelemetry will automatically instrument the signing operations. You can configure the telemetry collection using standard OpenTelemetry environment variables:
```bash
# Configure OTLP endpoint (e.g., Jaeger, Zipkin)
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317
# Set service name for traces
export OTEL_SERVICE_NAME=model-signing
# Configure exporters (default: otlp)
export OTEL_TRACES_EXPORTER=otlp
export OTEL_METRICS_EXPORTER=none
```
Example with tracing enabled:
```bash
# Start Jaeger (example using Docker)
docker run -d --name jaeger \
-p 16686:16686 \
-p 4317:4317 \
jaegertracing/all-in-one:latest
# Sign with tracing
OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317 \
OTEL_SERVICE_NAME=model-signing \
OTEL_TRACES_EXPORTER=otlp \
OTEL_METRICS_EXPORTER=none \
model_signing sign bert-base-uncased
```
#### Logging Configuration
The `model-signing` CLI supports configurable logging levels to help with debugging and monitoring:
```bash
# Enable debug logging on the command line
model_signing --log-level debug sign bert-base-uncased
# Or using environment variable
MODEL_SIGNING_LOG_LEVEL=debug model_signing sign bert-base-uncased
```
Available log levels: `DEBUG`, `INFO`, `WARNING`, `ERROR`, `CRITICAL`
### Model Signing API
We offer an API which can be used in integrations with ML frameworks, ML
pipelins and ML model hubs libraries. The CLI wraps around the API.
The API is split into 3 main components:
- `model_signing.hashing`: responsible with generating a list of hashes for
every component of the model. A component could be a file, a file shard, a
tensor, etc., depending on the method used. We currently support only files
and file shards. The result of hashing is a manifest, a listing of hashes for
every object in the model.
- `model_signing.signing`: responsible with taking the manifest and generating a
signature, based on a signing configuration. The signing configuration can
select the method used to sign as well as the parameters.
- `model_signing.verifying`: responsible with taking a signature and verifying
it. If the cryptographic parts of the signature can be validated, the
verification layer would return an expanded manifest which can then be
compared agains a manifest obtained from hashing the existing model. If the
two manifest don't match then the model integrity was compromised and the
`model_signing` package detected that.
The first two of these components allows configurability but can also be used
directly, with a default configuration. The only difference is for the
verification component where we need to configure the verification method since
there are no sensible defaults that can be used.
The simplest way to generate a signature using Sigstore is:
```python
import model_signing
model_signing.signing.sign("bert-base-uncased", "model.sig")
```
This will run the same OIDC flow as when signing with Sigstore from the CLI.
We can use explicit configurations to configure more about the signing:
```python
import model_signing
model_signing.signing.Config().use_elliptic_key_signer(
private_key="key.priv"
).sign(
"finbert", "finbert.sig"
)
```
The same signing configuration can be used to sign multiple models:
```python
import model_signing
signing_config = model_signing.signing.Config().use_elliptic_key_signer(
private_key="key.priv"
)
for model in all_models:
signing_config.sign(model, f"{model}_sharded.sig")
```
Verification needs a configuration. To verify using Sigstore:
```python
import model_signing
model_signing.verifying.Config().use_sigstore_verifier(
identity=identity, oidc_issuer=oidc_provider
).verify("finbert", "finbert.sig")
```
The same verification configuration can be used to verify multiple models:
```python
import model_signing
verifying_config = model_signing.signing.Config().use_elliptic_key_verifier(
public_key="key.pub"
)
for model in all_models:
verifying_config.verify(model, f"{model}_sharded.sig")
```
Consult the
[official documentation](https://sigstore.github.io/model-transparency/model_signing.html)
for more details.
### Model Signing Format
For a diagram showing the model signing format as well as an explanation of the
layers, see the [model signing format](docs/model_signing_format.md) document.
## Contributing
Please see the [Contributor Guide](CONTRIBUTING.md) for more information.
[openssf]: https://openssf.org/
