CEA-LIST / adv-reid

Metric Adversarial Attacks and Defense

Vulnerability of Person Re-Identification Models to Metric Adversarial Attacks

This repository contains the reference source code for the paper Vulnerability of Person Re-Identification Models to Metric Adversarial Attacks (CVPRW'20). For more information, please check out our paper, the presentation or the associated blogpost.

As classification attacks are not applicable to re-identification, we propose different possible attacks on metric learning models depending on the number and type of guides available. Two particularly effective attacks stand out. To defend against these attacks, we adapt the adversarial training protocol for metric learning. Let us guide you !

Requirements

• python==3.6+
• torch==1.5.0
• torchvision==0.6.0
• advertorch==0.2.0

You can run pip install -r requirements.txt to install required packages or conda env create -f environment.yml to create a new environment with the required packages installed.

As we use Market1501 and DukeMTMC-reid datasets for our experiments, you must download them beforehand.

Training

To train a Classification model on dataset DukeMTMC-reid, run:

python train.py \
         <dataset_folder> \
         --dataset 'duke' \
         --lr 0.00035 \
         -wd 5e-4 \
         -n 60 \
         -f <checkpoint_name> \
         -b 32 \
         --classif \
         --pretrained 

For a Triplet model on dataset Market1501, run:

python train.py \
        <dataset_folder> \
        --dataset 'market' \
        --lr 0.0003 \
        -wd 0 \
        -n 100 \
        -f <checkpoint_name> \
        -b 72 \
        -ni 4 \
        -e 2048 \
        --pretrained \
        --triplet \
        --soft \
        --id_batch

By default, the checkpoints of the trained models are saved in the folder ./models/checkpoints/.

Evaluating

To evaluate a Classification model under attack, run:

python gen_adv.py \
        <checkpoints_folder> \
        <dataset_folder> \
        <checkpoint_name> \
        --dataset <dataset> \
        --classif \
        -a <attack> \
        --eps 5

Defending

Before training a defended model, run sort_dataset.py to create a sorted dataset from which the guides are sampled.
Then, to train a Classification model using GOAT with FNA, run:

python train.py \
        <dataset_folder> \
        --dataset <dataset> \
        --lr 0.00035 \
        -wd 5e-4 \
        -n 60 \
        -f <checkpoint_name> \
        -b 32 \
        --classif \
        --pretrained \
        --adv \
        --push \
        --pull

Pretrained Defended Models

You can download the weights of the pretrained defended models with GOAT here :

Training Dataset	Pushing guides	Pulling guides	download link
Duke	0 4	0 1	duke_goat_classif_sma.th duke_goat_classif.th
Market	0 4	0 1	market_goat_classif_sma.th market_goat_classif.th

License

This repository is released under the CeCILL license, a free software license adapted to both international and French legal matters that is fully compatible with the FSF's GNU/GPL license.

Citation

If you find this repository useful for your own work, please cite our paper:

@InProceedings{Bouniot_2020_CVPR_Workshops,
  author = {Bouniot, Quentin and Audigier, Romaric and Loesch, Angélique},
  title = {Vulnerability of Person Re-Identification Models to Metric Adversarial Attacks},
  booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  month = {June},
  year = {2020}
}