Two Heads are Actually Better than One: Towards Better Adversarial Robustness by Combining Transduction and Rejection

Preliminaries

The dependencies are specified in environment.yml and can be installed with conda env create -f environment.yml.

Running the Experiments

python run_experiments.py runs all experiments; alternatively, the ith experiment in experiments may be run with python run_experiments.py i. SLURM and Ray systems are currently supported.

Overview of the Code

• train.py: training code for inductive and transductive training.
• models/detectors.py: wrappers which transform classifiers into selective classifiers.
• utils/attack_losses.py: loss functions for adaptive attacks targeting selective classifiers.
• utils/attacks.py: inductive attacks.
• utils/transductive_attacks.py: adaptive attacks targeting transduction.
• utils/losses.py: loss functions for use in training.
• utils/evaluate.py: tools to evaluate models.
• experiments.py: specifies the the experiments to be run.
• run_experiments.py: code to run the experiments, targeting SLURM.
• experiment_setup.py: generates configurations for synthetic data, MNIST, and CIFAR-10.

Acknowledgements

Part of the code is based on TRADES, GMSA, and Active Adversarial Tests.