Go To NNT V1.2 Data Types

1 Definition 2 Functional Requirements 3 Syntax 4 Semantics 5    Conformance Testing 6     Performance Assessment

1 Definition

A Data Type stating where a watermark sits in a neural network model and by which method it was inserted. It is neither part of the model (ML Model Object) nor of the Payload: an AIM that inserts a watermark gives it, and an AIM that extracts a watermark may take it.

2 Functional Requirements

A Locator includes:

  1. The ID of the M-Instance and of the U-Environment to which it belongs.
  2. The ID of the Locator and its time.
  3. The watermarking method that inserted the watermark, named as its Implementer names it. An extraction AIM can retrieve the Payload only if its Implementation implements this method.
  4. Optionally, the tensors of the model that carry the watermark, identified by the names used in the model file, whose format the ML Model Qualifier states (for example ONNX or PyTorch). If they are omitted, the method decides which tensors carry the watermark.
  5. Optionally, the identifier of the secret key that the method uses.
  6. Data Exchange Metadata providing additional information on the Locator.

An extraction AIM that is not given a Locator looks for the watermark where its own method would put it.

The Locator never carries a secret key, only a reference to it. The key is kept in Storage (in the Secure Profile, in Secure Storage reached through the Security API) under a rule that names the AIMs that may read it. A key in Storage can be read only by AIMs of the Module of the Controller that holds it: where the extraction AIM runs under another Controller, the key is provided to it by means outside this specification.

3 Syntax

https://schemas.mpai.community/NNW1/V1.2/data/Locator.json

4 Semantics

Label Description
Header Locator Header – Standard “NNW-WLC-Vx.y”
MInstanceID Identifier of M-Instance.
UEnvironmentID Identifier of U-Environment.
LocatorID Identifier of the Locator.
LocatorTime Time of the Locator.
Method The watermarking method that inserted the watermark, named as its Implementer names it.
Targets The tensors of the model that carry the watermark, identified by the names used in the model file. If omitted, the method decides.
KeyReference The identifier of the secret key that the method uses. The key is never carried.
DataXMData Information about this Locator instance.
DescrMetadata Descriptive Metadata.

5 Conformance Testing

A Data instance Conforms with Locator (NNW-WLC) if:

  1. Its JSON Object validates against its JSON Schema.
  2. Any included JSON Object validates against its JSON Schema.
  3. All Data in the JSON Object:
    1. Have the specified Data Types.
    2. Conform with the Qualifiers signaled in their JSON Schemas.

6 Performance Assessment

Go To NNT V1.2 Data Types