(Informative)

Classification task

The NN watermarking state of the art studies consider the classification as a predilection task.

Within this task, the inference of a neural network belongs to a fix set of predefined classes.

To evaluate the impact of injecting a watermark in a classification NN:

  • Probability of false alarm: and Precision:
  • Probability of missed detection: and Recall:

As these measures are based on binary classification problem, for multiclass classifiers the average for all classes shall be computed.

Image/speech processing tasks

The inference of a neural network is a produced content. For example, a neural network for speech synthesis will return an artificial voice based on a text. Every qualitative/quantitative evaluation of a content can be use:

  • Image: PSNR, SSIM, NCC, in addition to subjective test (e.g. as specified by ITU)
  • Speech recognition: Word/Sentence error rate, Intent recognition rate, in addition to subjective test (e.g. as specified by ITU)

Image semantic segmentation

The inference of a neural network is a semantic-labelled. To evaluate this method, we propose:

  • Precision: tp/(tp+fp)
  • Recall tp/(tp+fn)
  • Intersection over Union (Area of overlap)/(Area of Union)