Bai-YT / AdaptiveSmoothing

Implementation of the paper "Improving the Accuracy-Robustness Trade-off of Classifiers via Adaptive Smoothing".

Improving the Accuracy-Robustness Trade-off of Classifiers via Adaptive Smoothing

This repository is the code base for the preprint paper Improving the Accuracy-Robustness Trade-off of Classifiers via Adaptive Smoothing.

We publically share two CIFAR-100 models that aim to defend the $\ell_\infty$ attack. Each of the proposed models rely on an accurate base classifier, a robust base classifier, and an optional "mixing network". The two proposed models share the same accurate base classifier but use two different robust base models and mixing networks.

Running RobustBench to replicate the results

Running the RobustBench benchmark should only require pytorch, torchvision, numpy, click, and robustbench packages.

Make a directory <YOUR_MODEL_ROOT_DIR> at a desired path to store the model checkpoints. Then, download the following models:

• Accurate base classifier: Big Transfer (BiT) ResNet-152 model finetuned on CIFAR-100 -- download
• Robust base classifier 1: WideResNet-70-16 model from this repo -- download and rename as cifar100_linf_edm_wrn70-16.pt.
• Robust base classifier 2: WideResNet-70-16 model from this repo -- download and rename as cifar100_linf_trades_wrn70-16.pt.
• Mixing network to be coupled with robust base classifier 1 -- download
• Mixing network to be coupled with robust base classifier 2 -- download

Now, organize <YOUR_MODEL_ROOT_DIR> following the structure below:

<YOUR_MODEL_ROOT_DIR>
│
└───Base
│   │   cifar100_linf_edm_wrn70-16.pt
│   │   cifar100_linf_trades_wrn70-16.pt
│   │   cifar100_bit_rn152.tar
│   
└───CompModel
    │   cifar-100_edm_best.pt
    │   cifar-100_trades_best.pt

To benchmark existing models with RobustBench, run the following:

python run_robustbench.py --root_dir <YOUR_MODEL_ROOT_DIR> --model_name {edm,trades}

Note that while the base classifiers may require additional (collected or synthesized) training data, the provided mixing networks were only trained on CIFAR-100 training data.

Training a new model

To train a new model with the provided code, install the full environment. We require the following packages: pytorch torchvision tensorboard pytorch_warmup numpy scipy matplotlib jupyter notebook ipykernel ipywidgets tqdm click PyYAML.

To train, run the following:

python run.py --training --config configs/xxx.yaml

To evaluate, run the following:

python run.py --eval --config configs/xxx.yaml