Towards Query-Efficient Black-Box Adversary with Zeroth-Order Natural Gradient Descent

Authors: Pu Zhao, Pin-yu Chen, Siyue Wang, Xue Lin6909-6916

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

Reproducibility Variable Result LLM Response
Research Type Experimental The empirical evaluations on image classification datasets demonstrate that ZO-NGD can obtain significantly lower model query complexities compared with state-of-the-art attack methods.
Researcher Affiliation Collaboration 1Northeastern University, Boston, MA 02115 2IBM Research, Yorktown Heights, NY 10598
Pseudocode Yes Algorithm 1 Framework of ZO-NGD. Require: The legitimate image x; the correct label t; the model to be queried; the learning rate λ; the sampling step size μ; Ensure: Adversarial perturbation δ; 1: initialize δ0 with all zeros; 2: for k = 0, ..., K do 3: Query the model with δk and obtain the probability p(t|x, δk) := p(t|x + δk); 4: for j = 1, ..., R do 5: Generate a random direction vector uj drawn from a uniform distribution over the surface of a unit sphere; 6: Query the model with x + δk + μuj and obtain p(t|x, δk + μuj); 7: end for 8: Estimate the gradients of the loss function ˆ f(δk) according to Eq. (16); 9: Estimate the gradients of the log-likelihood function ˆ log p(t |x, δk ) according to Eq. (17); 10: Compute the FIM F according to Eq. (18) and perform the nature gradient update as shown in Eq. (19). 11: end for
Open Source Code Yes 1The code and appendix are available at https://github.com/LinLabNEU/ZO-NGD-blackbox.
Open Datasets Yes We compare ZO-NGD with various attack methods on three image classification datasets, MNIST (Lecun et al. 1998), CIFAR-10 (Krizhevsky and Hinton 2009) and Image Net (Deng et al. 2009).
Dataset Splits No The paper mentions selecting images from 'test datasets' for evaluation, and using a 'pretrained Inception v3 network' for ImageNet, but does not explicitly provide training/validation/test dataset splits (e.g., percentages or counts) for their own model training or for the attack evaluation.
Hardware Specification Yes All experiments are performed on machines with NVIDIA GTX 1080 TI GPUs.
Software Dependencies No The paper does not specify software dependencies with version numbers (e.g., Python version, specific library versions like TensorFlow, PyTorch, or Scikit-learn).
Experiment Setup Yes In the proposed ZO-NGD method, the sampling number R in the random gradient estimation as defined in Eq. (16) and (17) is set to 40. ϵ is set to 0.4 for MNIST and 0.2 for CIFAR-10 or Image Net. In Eq. (16) and (17), we set μ = 1 for three datasets. γ is set to 0.01.