Minimally distorted Adversarial Examples with a Fast Adaptive Boundary Attack

Authors: Francesco Croce, Matthias Hein

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We run experiments on MNIST, CIFAR-10 (Krizhevsky et al., 2014) and Restricted Image Net (Tsipras et al., 2019). For each dataset we consider a normally trained model (plain) and two adversarially trained ones as in (Madry et al., 2018) wrt the l -norm (l -AT) and the l2-norm (l2-AT) (see supplementary material for details). [...] We compare the performance of FAB-attack1 to those of attacks representing the state-of-the-art in each norm [...]. We report the complete results in the supplementary material, while we summarize them in Table 1 (MNIST and CIFAR-10 aggregated, as we used the same attacks) and Table 2 (Restricted Image Net). Our FAB-attack achieves the best results in all statistics for every norm (with the only exception of "max diff. to best" in l ) on MNIST+CIFAR-10.
Researcher Affiliation Academia 1University of Tübingen, Germany. Correspondence to: F. Croce <francesco.croce@uni-tuebingen.de>.
Pseudocode Yes Algorithm 1 FAB-attack
Open Source Code Yes 1https://github.com/fra31/fab-attack
Open Datasets Yes We run experiments on MNIST, CIFAR-10 (Krizhevsky et al., 2014) and Restricted Image Net (Tsipras et al., 2019).
Dataset Splits Yes In total we have 5 thresholds × 6 models = 30 cases for each of the 3 norms. *Note that for FAB-10 (i.e. with 10 restarts) the "# best" is computed excluding the results of FAB-100. (on the frst 1000 points for l and l1, 500 for l2, of the test sets)." (Table 1 caption) and "on the frst 500 points of the validation set" (Table 2 caption).
Hardware Specification No No specific hardware details (like exact GPU/CPU models, processor types, or memory amounts) used for experiments were provided.
Software Dependencies No No specific software dependencies with version numbers (e.g., library or solver names with version numbers like Python 3.8, PyTorch 1.9) were provided.
Experiment Setup Yes For FAB-attack we use always β = 0.9 and on MNIST and CIFAR-10: αmax = 0.1, η = 1.05 and on Restricted Image Net: αmax = 0.05, η = 1.3. These parameters are the same for all norms. ... As a result we use for PGD wrt l step size ϵ/10 and the direction is the sign of the cross entropy loss, for PGD wrt l2 we do a step in the direction of the l2-normalized gradient with step size ϵ/4, for PGD wrt l1 we use the gradient step suggested in (Tramèr & Boneh, 2019) (with sparsity levels of 1% for MNIST and 10% for CIFAR-10 and Restricted Image Net) with step size ϵ/2.