MixedNUTS: Training-Free Accuracy-Robustness Balance via Nonlinearly Mixed Classifiers

This is the official code implementation of the preprint paper
MixedNUTS: Training-Free Accuracy-Robustness Balance via Nonlinearly Mixed Classifiers
by Yatong Bai, Mo Zhou, Vishal M. Patel, and Somayeh Sojoudi.

TL;DR: MixedNUTS balances clean data classification accuracy and adversarial robustness without additional training via a mixed classifier with nonlinear base model logit transformations.

MixedNUTS has been accepted to Transactions on Machine Learning Research (TMLR).

Getting Started

Model Checkpoints

MixedNUTS is a training-free method that has no additional neural network components other than its base classifiers.

All robust base classifiers used in the main results of our paper are available on RobustBench and can be downloaded automatically via the RobustBench API.

Here, we provide the download links to the standard base classifiers used in the main results.

Dataset	Link
CIFAR-10	Download
CIFAR-100	Download
ImageNet	Download

After downloading the accurate base classifiers, create a base_models directory and organize as follows:

base_models
│
└───cifar10
│   └───cifar10_std_rn152.pt
│   
└───cifar100
    └───cifar100_std_rn152.pt
│   
└───imagenet
    └───imagenet_std_convnext_v2-l_224.pt

The CIFAR-10 and -100 accurate base classifiers are fine-tuned from BiT checkpoints. The ImageNet accurate base classifier is from the ConvNeXt-V2 repository.

Environment

Run the following to install the environment:

conda env create -f environment.yml

There are two additional pip packages that are not present in pypi and need to be manually installed via the following:

conda activate nlmc
pip install git+https://github.com/fra31/auto-attack
pip install git+https://github.com/RobustBench/robustbench.git

Replicating Our Results

CIFAR-10:

python run_robustbench.py --root_dir base_models --dataset_name cifar10 \
    --rob_model_name Peng2023Robust --std_model_arch rn152 --map_type best \
    --adaptive --n_examples 10000 --batch_size_per_gpu 40  --disable_nonlin_for_grad

CIFAR-100:

python run_robustbench.py --root_dir base_models --dataset_name cifar100 \
    --rob_model_name Wang2023Better_WRN-70-16 --std_model_arch rn152 --map_type best \
    --adaptive --n_examples 10000 --batch_size_per_gpu 40 --disable_nonlin_for_grad

ImageNet:

python run_robustbench.py --root_dir base_models --dataset_name imagenet \
    --rob_model_name Liu2023Comprehensive_Swin-L --std_model_arch convnext_v2-l_224 --map_type best \
    --adaptive --n_examples 5000 --batch_size_per_gpu 20 --disable_nonlin_for_grad

Building MixedNUTS with Your Base Classifiers

Please refer to scripts.sh for the workflow of constructing MixedNUTS.

---

---

Citing our work (BibTeX)

@article{MixedNUTS,
  title={MixedNUTS: Training-Free Accuracy-Robustness Balance via Nonlinearly Mixed Classifiers},
  author={Bai, Yatong and Zhou, Mo and Patel, Vishal M. and Sojoudi, Somayeh},
  journal={Transactions on Machine Learning Research},
  year={2024}
}

Third-party Code

adaptive_autoattack.py, autopgd_base.py, and fab_pt.py in adaptive_autoattack are modified based on AutoAttack.

robust_bench.py is modified based on RobustBench.