Query-efficient Meta Attack to Deep Neural Networks

Authors: Jiawei Du, Hu Zhang, Joey Tianyi Zhou, Yi Yang, Jiashi Feng

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on MNIST, CIFAR10 and tiny-Imagenet demonstrate that our meta-attack method can remarkably reduce the number of model queries without sacrificing the attack performance.
Researcher Affiliation Academia 1Dept. ECE, National University of Singapore, Singapore 2Re LER, University of Technology Sydney, Australia 3Institute of High performance Computing, A*STAR, Singapore
Pseudocode Yes Algorithm 1 Meta Attacker Training
Open Source Code Yes The code of our work is available at https://github.com/dydjw9/Meta Attack_ICLR2020/.
Open Datasets Yes We evaluate the attack performance on MNIST (Le Cun, 1998) for handwritten digit recognition, CIFAR10 (Krizhevsky & Hinton, 2009) and tiny-Imagenet (Russakovsky et al., 2015) for object classification.
Dataset Splits No The paper states, 'We use 10000 randomly selected images from the training set to train the meta-attackers in three datasets. The proportion of the selected images to the whole training set are 16%, 20%, and 10% respectively.' and 'we randomly select 1000 images from each dataset as test images.' However, it does not explicitly describe a validation set or its split.
Hardware Specification No The paper does not explicitly describe the hardware used to run its experiments, such as specific GPU or CPU models.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., 'Python 3.8, PyTorch 1.9').
Experiment Setup Yes Meta-training Details For all the experiments, we use the same architecture for the meta attacker A as shown in Table 6. We use Reptile (Nichol et al., 2018) with 0.01 learning rate to train meta attackers. [...] Fine-tuning parameters are set as m = 5 for MNIST and CIFAR10, and m = 3 for tiny-Imagenet. Top q = 128 coordinates are selected as part coordinates for attacker fine-tuning and model attacking on MNIST; and q = 500 on CIFAR10 and tiny-Imagenet.