evaluating-adaptive-test-time-defenses

evaluating-adaptive-test-time-defenses copied to clipboard

Evaluating the Adversarial Robustness of Adaptive Test-time Defenses

Francesco Croce*, Sven Gowal*, Thomas Brunner*, Evan Shelhamer*, Matthias Hein, Taylan Cemgil
https://arxiv.org/abs/2202.13711

Case study

We evaluate the following defenses:

• yoon_2021: Adversarial Purification with Score-based Generative Models
• hwang_2021: AID-purifier: A light auxiliary network for boosting adversarial defense
• wu_2021: Attacking Adversarial Attacks as A Defense
• shi_2020: Online Adversarial Purification based on Self-Supervision
• kang_2021: Stable Neural ODE with Lyapunov-Stable Equilibrium Points for Defending against Adversarial Attacks
• mao_2021: Adversarial Attacks are Reversible with Natural Supervision
• qian_2021: Improving Model Robustness with Latent Distribution Locally and Globally
• alfarra_2021: Combating Adversaries with Anti-Adversaries
• chen_2021: Towards Robust Neural Networks via Close-loop Control

Some folders have a single Python notebook while other contain more involved code. As a result, such folders will contain a run_eval.sh with the commands to run the evaluations or an explanatory README.md file. The pre-trained models have to be downloaded following the indications in the corresponding folders and papers, together with the details provided in the appendix of our paper.