This project aims to verify and analyze some of the properties of the neural networks, like Whether a network classifies the image correctly, incorrectly, or fails to classify when a small perturbation is applied to the image.
In this project, we use an already implemented analyzer ETH Robustness Analyzer for Neural Networks (ERAN) developed at the SRI Lab, Department of Computer Science, ETH Zurich for the verification of networks trained on MNIST and CIFAR-10 datasets. In this project, We use only a part of the ERAN, i.e DeepPoly which is used to verify the properties of a deep neural network over the perturbed inputs.
Follow the instructions to install ERAN from official git repo.