FakeSal

Overview

FakeSal is a saliency algorithm that does stuff as reported in FakeSal: A Fake Saliency Algorithm (2021)

To get started using FakeSal, install fake-package

pip install fake-package

To enable FakeSal, adjust the fake-config.yaml file as follows:

debug-options:
  use-saliency: true
  algorithm: fakesal

Intended Use

When to use FakeSal

  • Don’t need a black-box algorithm
  • Runtime performance is important

Model/Data

The input to FakeSal is a 224x224 raster image, a class label, and a handle to a CNN-based model and the output is a 224x224 full-color heatmap.

Limitations

As a whitebox model, FakeSal requires knowledge/access to the model we seek to explain. FakeSal is not robust to adversarial perturbations, and it has been shown FakeSal can produce extremely similar saliency maps in situations where the final predictions differ.

References

@InProceedings{fake_sal_2021,
  author = {Keresearcher, F. A. and Author, Invented and Menon, Nitesh},
  title = {FakeSal: A Fake Saliency Algorithm (2021)},
  booktitle = {arXiv},
  month = {December},
  year = {2021}
}