A two layer 01 loss neural network and its adversarial and corruption robustness

Introduction

This repository contains source code to reproduce experiment results in the paper.

Directory

blackbox_attack contains the black box attack substitute models' definition.
core contains scd01 and scd01mlp for binary and one-vs-all version for multi-class.
main contains training scripts for scd and its variations, svm, mlp-logistic, training, black-box attack, corruption and perturbation evaluation.
tools contains args definition, data set loading and other tools will be used in training scrips.
whitebox_attack contains script for white-box attack experiments.

Python v3.6.8

Pytorch v1.3.1

Numpy v1.17.4

Pandas v1.0.3

Argparse v1.1

Sklearn v0.21.2