Submission + - Verification of Neural Networks via Linear Programming (github.com)
An anonymous reader writes: A step-by-step tutorial on how to check the robustness of a neural network and how to systematically create adversarial examples using linear programming. An adversarial image is an image that is altered just enough for a neural network to misclassify it. The original image and the altered (adversarial) one are indistinguishable to the human eye. The tutorial includes Python code on an Jupyter notebook. The goal of the presentation below is to be simple yet precise. It is meant to illustrate a technique rather than to provide a practical implementation.