Go to AI Workflows

1 Functions 2 Reference Architecture 3 I/O Data
4 Functions of AI Modules 5 I/O Data of AI Modules 6 AIW, AIMs, and JSON Metadata
7 Reference Software 8 Conformance Texting 9 Performance Assessment

1        Functions

The No Training Imperceptibility (NNW-NTI) receives a dataset separated in training and testing, the payload and provides both inference for unwatermarked and watermarked AIM allowing the imperceptibility evaluation.

2        Reference Model

Figure 1 specifies the No Training Imperceptibility (NNW-NTI) Reference Model including the input/output data, the AIMs, and the data exchanged between and among the AIMs.

Figure 1 – Reference Model of No Training Imperceptibility (NNW-NTI)

The operation of No Training Imperceptibility (NNW-NTI) develops in the following way:

  1. A user provides
    1. The training and testing datasets
    2. The Payload
  2. The machine provides
    1. The unwatermarked inference
    2. The watermarking inference

3         I/O Data

The input and output data of the No Training Imperceptibility (NNW-NTI) Use Case are:

Table 1 – I/O Data of No Training Imperceptibility (NNW-NTI)

Input Descriptions
Training dataset The dataset used to train the AIM.
Testing dataset The part of the dataset unseen during training.
Payload The information to be inserted in the AIM.
Output Descriptions
Unwatermarked inference The output of the unwatermarked AIM on the testing dataset.
Watermarked inference The output of the watermarked AIM on the testing dataset.

4        Functions of AI Modules

Table 2 provides the functions of the No Training Imperceptibility (NNW-NTI) Use Case.

Table 2 – Functions of AI Modules of No Training Imperceptibility (NNW-NTI)

AIM Function
AIM Trainer Trains the AIM using training dataset.
NTI Watermark Embedder Produces watermarked parameters.
Unwatermarked AIM Produces an inference using unwatermarked parameters and testing dataset.
Watermarked AIM Produces an inference using watermarked parameters and testing dataset.

5         I/O Data of AI Modules

The AI Modules of No Training Imperceptibility (NNW-NTI) are given in Table 3.

Table 3 – AI Modules of No Training Imperceptibility (NNW-NTI)

AIM Receives Produces
AIM Trainer Training dataset Unwatermarked parameters
NTI Watermark Embedder 1.     Unwatermarked parameters

2.     Payload

 Watermarked parameter
Unwatermarked AIM 1.     Unwatermarked parameters

2.     Testing dataset

 Unwatermarked inference
Watermarked AIM 1.     Watermarked parameters

2.     Testing dataset

 Watermarked inference

6        AIW, AIMs, and JSON Metadata

Table 4 provides the links to the AIW and AIM specifications and to the JSON syntaxes. AIMs/1 indicates that the column contains Composite AIMs and AIMs indicates that the column contains their Basic AIMs.

Table 4 – AIW, AIMs, and JSON Metadata

AIW AIM Name JSON
NNW-NTI No Training Imperceptibility X
NNW-MTR Module Trainer X
NNW-NWE NTI Watermark Embedder X
NNW-UWM Unwatermarked Module X
NNW-WMM Watermarked Module X

7        Reference Software

7.1       Disclaimers

  1. This NNW-NTI Reference Software Implementation is released with the BSD-3-Clause licence.
  2. The purpose of this Reference Software is to demonstrate a working Implementation of NNW-NTI, not to provide a ready-to-use product.
  3. MPAI disclaims the suitability of the Software for any other purposes and does not guarantee that it is secure.
  4. Use of this Reference Software may require acceptance of licences from the respective repositories. Users shall verify that they have the right to use any third-party software required by this Reference Software.

7.2        Guide to the NNW-NTI code

Use of this AI Workflow is for developers who are familiar with Python and PyTorch libraries,

The imperceptibility.py code allow a User to evaluate the imperceptibility of a watermarking method on the image classification task:

  1. The watermarking method is implemented as a Python Class
  2. The AIM Trainer uses CIFAR10 dataset to trains torchvision.models
  3. The fulltest function allow to provide the inference of both unwatermarked and watermarked AIM

The NNW-NTI Reference Software is found at the gitlab site. It contains:

  1. The python code implementing the AIW.
  2. The required libraries are: pytorch, tqdm

8        Conformance Testing

9        Performance Assessment

Go to AI Workflows