Adversarially Robust Models may not Transfer Better

This repo provides the implementation of the paper Adversarially Robust Models may not Transfer Better: Sufficient Conditions for Domain Transferability from the View of Regularization.

Prerequisites

The code successfully run on Python 3.6 + PyTorch 1.8.1. The PyTorch package needs to be installed based on different hardware environments. Other packages can be installed by pip install -r requirements.txt.

Training

python {train,train_aug,train_lastot,train_reg}.py

• The hyper-parameters are specified within each script.

Evaluation

python {eval_attack,eval_transfer}.py

• The hyper-parameters are specified within each script.