Provably Robust Cost-Sensitive Learning via Randomized Smoothing

The goal of this project:

• We study whether randomized smoothing, a scalable certification framework, can be leveraged to certify and train for cost-sensitive robustness.
• To certify cost-sensitive robustness, we propose a new practical certification pipeline to produce tight robustness guarantees for any given cost matrix.
• To train smoothed classifiers that are optimized for cost-sensitive robustness, we propose fine-grained certified radius optimization schemes.

Installation & Usage

• Install PyTorch

  pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116
  

• Install Other Dependencies

  pip install scipy pandas statsmodels matplotlib seaborn
  

• Certification for overall robustness

  python certify.py --dataset cifar10 --base_classifier checkpoint_dir --sigma 0.5 --outfile outfile --certify overall --type macer
  

• Certification for cost-sensitive robustness

  python certify.py --dataset cifar10 --base_classifier checkpoint_dir --sigma 0.5 --outfile outfile --certify sensitive --type macer
  

• Training for seedwise cost matrices

  python train.py --dataset cifar10 --version v0 --ckptdir ckpt --lbd1 3  --sigma 0.5 --seed_type 3
  

• Training for pairwise cost matrices

  python train.py --dataset cifar10 --version v1 --ckptdir ckpt --lbd1 3  --sigma 0.5 --seed_type 3 --target_type single
  

What is in this repository?

• CS-RS, including:
  • train.py: implements detailed training pipeline
  • macer.py, pair_macer.py: main functions for training provably robust cost-sensitive classifiers for seedwise and pairwise cost matrices
  • certify: implements detailed certification pipeline
  • core.py: implements detailed practical certification algorithm for cost-sensitive robustness
  • train_gaussian_R.py, train_salman_R.py,train_smoothmix_R.py: implements the reweighting adaptation for three baselines: standard Gaussian, SmoothAdv and SmoothMix
  • analyze.py: functions to draw the cost-sensitive robustness v.s. certified radius curves