A Few Seconds Can Change Everything: Fast Decision-based Attacks against DNNs

Authors: Ningping Mou, Baolin Zheng, Qian Wang, Yunjie Ge, Binqing Guo

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on three datasets demonstrate that Fast Drop can escape the detection of the stateof-the-art (SOTA) black-box defenses and reduce the number of queries by 13 133 under the same level of perturbations compared with the SOTA attacks.
Researcher Affiliation Academia Wuhan University, Wuhan, China {ningpingmou, baolinzheng, qianwang, yunjiege, binqingguo}@whu.edu.cn
Pseudocode Yes Algorithm 1 Orderly Frequency Dropping Input: Original input (x, y), target model f Output: Adversarial input (x , y ) 1: Famplitude, Fphrase FFT(x). 2: {b1, b2, ..., bn} split Famplitude into blocks. 3: {s1, s2, ..., sn} Sort({b1, b2, ..., bn}). 4: for i = 1 : n do 5: si 0. 6: x IFFT(Famplitude, Fphrase). 7: if f(x ) = y then 8: x x . 9: break. 10: end if 11: end for 12: return x
Open Source Code No The paper does not provide an explicit statement about the release of its own source code, nor does it include a link to a code repository for the methodology described.
Open Datasets Yes We use commonly-used Image Net [Deng et al., 2009], Flowers-102 [Nilsback and Zisserman, 2008], and STL-10 [Coates et al., 2011] as our datasets.
Dataset Splits No The paper mentions Image Net, Flowers-102, and STL-10 as datasets and states that 1000 images are randomly selected from each class. However, it does not specify explicit train/validation/test splits for their experiments.
Hardware Specification Yes Moreover, when conducting experiments of Res Net50 on a Ge Force RTX 3080, BOA-1000 needs 207.15s to finish an attack of an image, while Fast Drop only needs 0.90s.
Software Dependencies No The paper mentions general tools and models like ResNet50 and MobileNetV3 (implying deep learning frameworks like PyTorch or TensorFlow), but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes To make fair comparisons, we carefully tune the hyper-parameters of these methods to achieve better results, and show the least queries under the l2 constraint. The paper also includes Appendix A titled "Hyperparameter Analysis" which details analysis of the "order of the sorted blocks", "non-zero modification", and "threshold of OFD" for Fast Drop.