Skip to content

Latest commit

 

History

History
50 lines (41 loc) · 1.68 KB

README.md

File metadata and controls

50 lines (41 loc) · 1.68 KB
Raw

README

Official repository for Interpreting Vulnerabilities of Multi-Instance Learning to Adversarial Perturbations.
Any question can contact with inki.yinji@gmail.com
My home pages:

Introduction

The code implements two MIL attackers: MI-CAP and MI-UAP

If the input is a bag with images, we annotate the file names as MI-CAP2D and MI-UAP2D
If the input is a bag with vectors, we annotate the file names as MI-CAP and MI-UAP

How to use

For ShanghaiTech and UCF-Crime data sets, just run MI-CAP and MI-UAP
For MNIST, CIFAR10, and STL10, just run MI-CAP2D and MI-UAP2D

For experimental parameters:

  • xi: The magnitude of perturbation
    • For ShanghaiTech and UCF-crime: the default setting is 0.01
    • For images: 0.2
  • mode: The computation mode for gradient, "ave" or "att"
  • net_type: The attacked network, the choice includes:
    1. ab: ABMIL
    2. ga: GAMIL
    3. la: LAMIL
    4. ds: DSMIL
    5. ma: MAMIL
  • data_type:
    • For MI-CAP and MI-UAP: "shanghai" or "ucf";
    • For MI_CAP2D and MI-UAP2D: "mnist", "cifar10", "stl10"

Citation

You can cite our paper as:

@article{Zhang:2023:109725,
author		=	{Yu-Xuan Zhang and Hua Meng and Xue-Mei Cao and Zheng Chun Zhou and Mei Yang and Avik Ranjan Adhikary},
title		=	{Interpreting vulnerabilities of multi-instance learning to adversarial perturbations},
journal		=	{Pattern Recognition},
pages		=	{109725},
year		=	{2023},
url			=	{https://doi.org/10.1016/j.patcog.2023.109725}
}