Demystifying the Adversarial Robustness of Random Transformation Defenses

Authors: Chawin Sitawarin, Zachary J Golan-Strieb, David Wagner

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental First, we show that the BPDA attack (Athalye et al., 2018a) used in Ba RT s evaluation is ineffective and likely overestimates its robustness. We then attempt to construct the strongest possible RT defense through the informed selection of transformations and Bayesian optimization for tuning their parameters. Furthermore, we create the strongest possible attack to evaluate our RT defense. Our new attack vastly outperforms the baseline, reducing the accuracy by 83% compared to the 19% reduction by the commonly used Eo T attack (4.3 improvement). Our result indicates that the RT defense on Imagenette dataset (a ten-class subset of Image Net) is not robust against adversarial examples. Extending the study further, we use our new attack to adversarially train RT defense (called Adv RT), resulting in a large robustness gain. Code is available at https://github.com/wagnergroup/demystify-random-transform.
Researcher Affiliation Academia 1Department of Electrical Engineering and Computer Sciences, University of California, Berkeley, Berkeley CA, USA.
Pseudocode Yes Algorithm 1 Our best attack on RT defenses
Open Source Code Yes Code is available at https://github.com/wagnergroup/demystify-random-transform.
Open Datasets Yes Our experiments use two datasets: CIFAR-10 and Imagenette (Howard, 2021), a ten-class subset of Image Net.
Dataset Splits Yes All models are trained for 70 epochs, and we save the weights with the highest accuracy on the held-out validation data (which does not overlap with the training or test set).
Hardware Specification Yes one BO run still takes approximately two days to complete on two GPUs (Nvidia Ge Force GTX 1080 Ti).
Software Dependencies No The paper mentions software tools like "Ray Tune library" and a "Bayesian optimization tool implemented by Nogueira (2014)" but does not specify version numbers for any software dependencies.
Experiment Setup Yes In all of the experiments, we use a learning rate of 0.05, batch size of 128, and weight decay of 0.0005. We use cosine annealing schedule (Loshchilov & Hutter, 2017) for the learning rate with a period of 10 epochs which also doubles after every period. All models are trained for 70 epochs, and we save the weights with the highest accuracy on the held-out validation data (which does not overlap with the training or test set).