Boosting Ray Search Procedure of Hard-label Attacks with Transfer-based Priors

Authors: Chen Ma, Xinjie Xu, Shuyu Cheng, Qi Xuan

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on the Image Net and CIFAR-10 datasets show that our approach significantly outperforms 11 state-of-the-art methods in terms of query efficiency.
Researcher Affiliation Collaboration 1 Institute of Cyberspace Security, Zhejiang University of Technology, Hangzhou 310023, China 2 Binjiang Institute of Artificial Intelligence, ZJUT, Hangzhou 310056, China 3 JQ Investments, Shanghai 200122, China EMAIL, EMAIL, EMAIL
Pseudocode Yes Algorithm 1 Prior-Sign-OPT and Prior-OPT attack
Open Source Code Yes To further support reproducibility, we provide the complete attack code for our approach and all baseline methods at https://github.com/machanic/hard_label_attacks.
Open Datasets Yes All experiments are conducted on two datasets, i.e., CIFAR-10 (Krizhevsky & Hinton, 2009) and Image Net (Deng et al., 2009).
Dataset Splits Yes We randomly select 1,000 images from the validation sets for experiments.
Hardware Specification Yes The experiments of all methods are conducted using Py Torch 1.7.1 framework on NVIDIA V100 and A100 GPUs.
Software Dependencies Yes The experiments of all methods are conducted using Py Torch 1.7.1 framework on NVIDIA V100 and A100 GPUs.
Experiment Setup Yes Table 3: The hyperparameters of Prior-OPT and Prior-Sign-OPT. Dataset Hyperparameter Value q, total number of vectors for estimating a gradient, including priors and random vectors 200 the binary search s stopping threshold β 500 the number of iterations 1,000 gmax, the maximum gradient norm for the gradient clipping operation 0.1