SawanGurjar / DeepPoly-Verification_of_NN

To verify and analyze classification properties of the neural networks when a small perturbation is applied to the image from MNIST, CIFAR-10 (and both with prepared Blured-Image) dataset.

Verification and Robustness Analysis of Neural Networks using DeepPoly

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.

Installation

Follow the instructions to install ERAN from official git repo.