Data-Free Hard-Label Robustness Stealing Attack

Authors: Xiaojian Yuan, Kejiang Chen, Wen Huang, Jie Zhang, Weiming Zhang, Nenghai Yu

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experiments demonstrate the effectiveness of our method. The clone model achieves a clean accuracy of 77.86% and a robust accuracy of 39.51% against Auto Attack, which are only 4.71% and 8.40% lower than the target model on the CIFAR-10 dataset, significantly exceeding the baselines.
Researcher Affiliation Academia 1University of Science and Technology of China 2Nanyang Technological University
Pseudocode Yes The framework is illustrated in Fig. 2 and the training algorithm is in the appendix.
Open Source Code Yes Our code is available at: https://github.com/Lethe Sec/DFHL-RS-Attack.
Open Datasets Yes We consider two benchmark datasets commonly used in AT research (Madry et al. 2018; Zhang et al. 2019; Li et al. 2023b), CIFAR-10 and CIFAR-100 (Krizhevsky, Hinton et al. 2009), as target datasets.
Dataset Splits No The paper describes training on datasets and reports test performance, but it does not explicitly specify validation dataset splits (e.g., percentages or counts for a separate validation set).
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions optimizers like Adam and SGD but does not specify software versions for libraries (e.g., PyTorch version) or other dependencies.
Experiment Setup Yes For substitute data generation, we use Adam optimizer with β = (0.5, 0.999) and set the hyperparameter λ = 3 in Eq. (7). For CIFAR-10, we set learning rates ηG = 0.002, ηz = 0.01, number of iterations NG = 10 and the label smoothing factor is set to 0.2. For CIFAR-100, we set learning rates ηG = 0.005, ηz = 0.015, number of iterations NG = 15 and the label smoothing factor is set to 0.02. For training MC, we use SGD optimizer with an initial learning rate of 0.1, a momentum of 0.9 and a weight decay of 1e 4. For constructing HEE, the step size α in Eq. (3) is set to 0.03 and the number of iterations is set to 10. We set the iterations of the clone model NC = 500. The batch sizes for CIFAR-10 and CIFAR100 are set to B = 256 and B = 512, respectively. We apply a cosine decay learning rate schedule and the training epoch is E = 300.