Attacking deep networks with surrogate-based adversarial black-box methods is easy

Authors: Nicholas A. Lord, Romain Mueller, Luca Bertinetto

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

Reproducibility Variable Result LLM Response
Research Type Experimental 3 EXPERIMENTS, Experimental setup: We compare GFCS to the methods of Cheng et al. (2019); Tashiro et al. (2020); Yang et al. (2020) by designing an experimental framework covering the key aspects of the original experiments in the respective source works. Results: Table 1 reports attack success rates and median query counts, and Fig. 2 plots cumulative success counts against the maximum queries spent per example (CDFs, modulo normalisation).
Researcher Affiliation Industry Nicholas A. Lord, Romain Mueller & Luca Bertinetto www.five.ai {nick,romain.mueller,luca.bertinetto}@five.ai
Pseudocode Yes Algorithm 1 GFCS: Gradient First, Coimage Second
Open Source Code Yes Code is available at https://github.com/fiveai/GFCS. We accompany this submission with the code implementing the proposed GFCS method.
Open Datasets Yes We use each method to perform ℓ2-norm-constrained untargeted attacks against the same 2000 randomly chosen correctly classified ILSVRC2012 validation images per victim network. using CIFAR-10 as the dataset
Dataset Splits Yes We use each method to perform ℓ2-norm-constrained untargeted attacks against the same 2000 randomly chosen correctly classified ILSVRC2012 validation images per victim network. We chose 2.0 as the default value for our experiments by performing a small grid search over a held-out set (disjoint from the 2000 examples used in the experiments of the main paper).
Hardware Specification No The paper mentions training on 'GPUs' implicitly through the context of deep learning, but does not specify any particular models (e.g., NVIDIA A100, RTX 2080 Ti) or other specific hardware components used for running experiments.
Software Dependencies No All networks used are pretrained models available via Py Torch/torchvision.
Experiment Setup Yes A maximum query count of 10000 is set per example (beyond which failure is declared), and the ℓ2 bound (enforced using PGA) is set to the commonly chosen 0.001D, where D is the image dimension in the victim network s native input resolution. We chose 2.0 as the default value for our experiments by performing a small grid search over a held-out set...