harrywuhust2022 / Reg_IBP_ICCV2021

This is the official implementation of ICCV2021 submitted paper (paper number: 6975)

Reg-IBP

This is an official implementation of submitted ICCV2021 paper "Reg-IBP: Efficient and Scalable Neural Network Robustness Training via Interval Bound Propagation"

Installation

1. Install pytorch
2. Clone this repository

Data Setup

1.MNIST dataset 2.CIFAR10 dataset 3.TinyImageNet dataset 4.ShanghaiTech part A & B dataset

Verifiably train the proposed Reg-IBP:

1. python3 tiny.py # Tiny imageNet challenge

2. python3 IBP_big_CIFAR_eps_8_255.py # CIFAR-10 challenge

3. python3 MNIST.py # reproduce the MNIST results

4. python3 soft_train.py # MCNN verifiably train

For Reproducibility: Our Reg-IBP trained models (CIFAR, MNIST) are available at:

Baidu Disk: https://pan.baidu.com/s/1TZ8Ndqw6-6bG1bTihJP3ZA code: hary

Dropbox for models:

trained models for MNIST and CIFAR10 datasets:
(Dropbox)
https://www.dropbox.com/sh/hn14wlkvg1m75k2/AABqXGC3PBjLTyU7lbQtHLDVa?dl=0