Generate Multi-label Adversarial Examples by Linear Programming

Requirements

• python-3.6.6
• numpy-1.16.5
• torch-1.3.1
• torchnet-0.0.4
• torchvision-0.2.0
• tqdm-4.31.1
• mosek-9.1.7

Usage

• apply license of the package mosek on https://www.mosek.com/products/academic-licenses/ and move the license file 'mosek.lic' to folder /home/username/mosek/mosek.lic

• create new conda environment(the environment name is mlalp)

  conda env create -f mlalp_conda_env.yaml

• download the VOC2007 and VOC2012 dataset and move to the folder 'data/voc2007/VOCdevkit/VOC2012/' or 'data/voc2012/VOCdevkit/VOC2012/'

• download ML-GCN model and ML-LIW model from model download or you can train the model yourself

  cd ml_gcn_model or cd ml_liw_model
  python train.py

• move the model to folder 'checkpoint/mlgcn' or 'checkpoint/mlliw'

• go the code folder

  cd src

• run attack

  python demo_mlgcn_voc2007.py --adv_batch_size=10 --adv_method='mla_lp' --target_type='hide_single'
  python demo_mlgcn_voc2012.py --adv_batch_size=10 --adv_method='mla_lp' --target_type='hide_single'
  python demo_mlliw_voc2007.py --adv_batch_size=10 --adv_method='mla_lp' --target_type='hide_single'
  python demo_mlliw_voc2012.py --adv_batch_size=10 --adv_method='mla_lp' --target_type='hide_single'

• test attack performance

Adversarial Data and Model

Thanks

• https://github.com/Megvii-Nanjing/ML-GCN