Awesome

AdvOrder: Practical Relative Order Attack

Paper Title: Practical Relative Order Attack in Deep Ranking, ICCV'2021
Preprint Link: https://arxiv.org/abs/2103.05248
Poster: https://github.com/cdluminate/advorder/blob/main/assets/advorder-poster.pdf
Slides: https://github.com/cdluminate/advorder/blob/main/assets/slides.pdf
Bilibili: https://www.bilibili.com/video/BV1Af4y1J7PD/
Youtube: https://www.youtube.com/watch?v=1fEDNPACpAA

General Usage

Step #	Description	Command
1	Download the datasets	`python3 Download.py`
2.1	Train network on fasion	`python3 Train.py -D cuda -M faC_c2f2`
2.2	Train network on sop	`python3 Train.py -D cuda -M sopE_res18`
3	Conduct white-box attack experiments	`bash bin/wtable`
4.1.1	(black box) Rand on fasion	`bash bin/farandsearch 5 50 4`
4.1.2	(black box) Beta on fasion	`bash bin/fabatk5 50 4`
4.1.3	(black box) PSO on fasion	`bash bin/fapso 5 50 4`
4.1.4	(black box) NES on fasion	`bash bin/fanes 5 50 4`
4.1.5	(black box) SPSA on fasion	`bash bin/faspsa 5 50 4`
4.2.1	(black box) SPSA on SOP	`bash bin/randsearch 5 50 4`
4.2.2	(black box) SPSA on SOP	`bash bin/batk 5 50 4`
4.2.3	(black box) SPSA on SOP	`bash bin/pso 5 50 4`
4.2.4	(black box) SPSA on SOP	`bash bin/nes 5 50 4`
4.2.5	(black box) SPSA on SOP	`bash bin/spsa 5 50 4`

The syntax of the bin/* black-box attack commands is bin/<algorithm> k N varepsilon*255. This project is forked from the codebase of my ECCV 2020 work "Adversarial Ranking Attack and Defense". And the command for training defensive model is more or less the same. Some functionalities used by ECCV 2020 but not used in the ICCV 2021 paper are simply invalidated with, e.g., raise NotImplementedError.

Hint: export USE_CPP_KERNEL=1 can significantly speed up the black-box attack experiments.

Hint: export USE_RUST_KERNEL=1 will use the Rust implementation of SRC which is even faster than the C++ one. Requires rustc.

Hint: Enabling Adam optimizer for SPSA (export SS_ADAM) may slightly boost the performance.

Detailed File Descriptions

.
├── Attack.py             | entrance script for white-box order attack
├── bin                   | collection of shortcut scripts
│   ├── batk              |   Beta on sop
│   ├── batkparam         |   Beta parameter search example
│   ├── batktable         |   Beta batched experiments
│   ├── fabatk            |   Beta on fashion
│   ├── fanes             |   NES  on fashion
│   ├── fapso             |   PSO  on fashion
│   ├── fapsops           |   PSO  parameter search example
│   ├── farandsearch      |   Rand on fashion
│   ├── faspsa            |   SPSA on fashion
│   ├── nes               |   NES  on sop
│   ├── nesparam          |   NES  parameter search example
│   ├── nestable          |   NES  batched experiments
│   ├── nodrtable         |   Ablation: no dimension reduction
│   ├── pso               |   PSO  on sop
│   ├── psoparam          |   PSO  parameter search example
│   ├── psotable          |   PSO  batched experiments
│   ├── qbudcurve.py      |   plot a figure in appendix
│   ├── randsearch        |   Rand on sop
│   ├── randsearchtable   |   Rand batched experiments
│   ├── spsa              |   SPSA on sop
│   ├── spsaparam         |   SPSA parameter search example
│   ├── spsatable         |   SPSA batched experiments
│   ├── wloss.py          |   plot a figure in main manuscript
│   ├── wloss-sop.py      |   plot a figure in appendix
│   └── wtable            |   white-box batched experiments
├── BlackOA.py            | entrance script for black-box order attack
├── config.yml            | configuration file for models and attacks
├── display.py            | display query results from SnapShop
├── _download.py          | helper utility used for download files
├── Download.py           | download the fashion minst dataset
├── lib                   | core algorithms
│   ├── common.py         |   white-box order attack implementation
│   ├── datasets          |   dataset abstractions
│   │   ├── fashion.py    |     fashion-mnist dataset
│   │   ├── __init__.py   |     python file
│   │   └── sop.py        |     stanford-online-products
│   ├── faC_c2f2.py       |   fashion c2f2 network with cosine metric
│   ├── faC_lenet.py      |   fashion lenet with cosine metric
│   ├── faC_res18.py      |   fashion resnet18 
│   ├── faE_c2f2.py       |   fashion c2f2 network with euclidean metric
│   ├── __init__.py       |   python file
│   ├── rankingmodel.py   |   abstract class
│   ├── reorder.py        |   black-box order attack implementation
│   ├── snapshop.py       |   snapshop client and abstraction
│   ├── sopE_res18.py     |   sop resnet18 with euclidean metric
│   ├── sopE_res50.py     |   sop resnet50 with euclidean metric
│   ├── srckernel_py.py   |   Primitive SRC implementation in python. (slow) 
│   ├── _srckernel.cc     |   SRC function in C++ (moderate speed)
│   ├── srckernel_cc.py   |   python wrapper for the C++ SRC function
│   ├── srck/*            |   Rust implementation of the SRC function. (fast)
│   ├── srckernel_rs.py   |   python wrapper for the Rust SRC function
│   ├── test_srckernel.py |   tester of the C++ SRC function
│   └── utils.py          |   miscellaneous
├── poc                   |   miscellaneous
│   └── taumap.py         |   miscellaneous
├── PracticalOA.py        | entrance script for practical order attack
├── Train.py              | entrance script for training ranking model
└── visrow.py             | visualization helper for snapshop attack

Software Version

Python  3.8.3
PyTorch 1.7.0
Numpy   1.18.5
Scipy   1.5.0
CUDA    11
Cargo   1.45.0
Rustc   1.48.0
Linux   5.10

Bibtex

@InProceedings{advorder,
    author    = {Zhou, Mo and Wang, Le and Niu, Zhenxing and Zhang, Qilin and Xu, Yinghui and Zheng, Nanning and Hua, Gang},
    title     = {Practical Relative Order Attack in Deep Ranking},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {16413-16422}
}

License Info

Copyright (C) 2020-2021 Mo Zhou <cdluminate@gmail.com>
Released under the Apache-2.0 License.