Go to ToC

1. Technical Specification 2. Reference Software 3. Conformance Testing 4. Performance Assessment

1. Technical Specifications

Table 1 provides the links to the specifications and the JSON syntax of all AIMs specified by Technical Specification: Neural Network Watermarking (MPAI-NNW) – Neural Network Traceability (MPAI-NNT) V1.2. AI Modules in bold are Composite.

Table 1 – Specifications and JSON syntax of AIMs of NNW-NNT V1.2

Acronym Name JSON Acronym Name JSON
NNW-PWI Post-Training Watermark Insertion X NNW-DWI During-Training Watermark Insertion X
NNW-WEX Watermark Extraction X

2 Conformance testing

A Data instance of a Data Type Conforms with NNW-NNT V1.2 if the JSON Data validate against the relevant NNW-NNT V1.2 JSON Schema and if the Data Conforms with the relevant Data Qualifier, if present. NNW-NNT V1.2 does not provide a method for testing the Conformance of the Semantics of the Data instance.Conformance testing can be performed by a human using a JSON Validator to verify the Conformance of the syntax of JSON Data to the relevant JSON Schema; and, if the Data has a Qualifier, to verify that the syntax of the Data conforms with the relevant values in the Data Qualifier. Alternatively, Conformance testing can be performed by software implementing the steps above.

3. Reference Software

As a rule, MPAI provides Reference Software implementing the AI Modules released with the BSD-3-Clause licence and the following disclaimers:

  1. The purpose of the Reference Software is to provide a working Implementation of an AIM, not a ready-to-use product.
  2. MPAI disclaims the suitability of the Software for any other purposes than those of the NNW-NNT Standard, and does not guarantee that it offers the best performance and that it is secure.
  3. Users shall verify that they have the right to use any third-party software required by this Reference Software, e.g., by accepting the licences from third-party repositories.

4. Performance Assessment

Performance is a multidimensional entity because it can have various connotations. Therefore, the Performance Assessment Specification should provide methods to measure how well an AIM performs its function, using a metric that depends on the nature of the function, such as:

  1. Quality: Performance Assessment measures how well an AIM performs its function, using a metric that depends on the nature of the function, e.g., how well a Watermark Extraction AIM retrieves the Payload inserted in a neural network.
  2. Bias: Performance Assessment measures the preference given by an AIM to certain elements, using a metric that depends on a bias related to certain attributes of the AIM.
  3. Legal compliance: Performance Assessment measures how well an AIM performs its function, using a metric that assesses its accordance with a certain legal standard.
  4. Ethical compliance: the Performance Assessment of an AIM can measure the compliance of an AIM to a target ethical standard.

Go to ToC