Imperceptible Adversarial Attack via Invertible Neural Networks

Authors: Zihan Chen, Ziyue Wang, Jun-Jie Huang, Wentao Zhao, Xiao Liu, Dejian Guan

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on CIFAR-10, CIFAR-100, and Image Net-1K demonstrate that the proposed Adv INN method can produce less imperceptible adversarial images than the state-of-the-art methods and Adv INN yields more robust adversarial examples with high confidence compared to other adversarial attacks.
Researcher Affiliation Academia College of Computer Science, National University of Defense Technology, Changsha, Hunan, China {chenzihan21, wangzy13, jjhuang, wtzhao, liuxiao13a, guandejian20}@nudt.edu.cn
Pseudocode No No pseudocode or clearly labeled algorithm blocks were found in the paper.
Open Source Code Yes Code is available at https://github.com/jjhuangcs/Adv INN.
Open Datasets Yes We evaluate the performance of the comparison methods on Image Net-1K dataset which contains 1000 images sampled from the Image Net-1K validation set (Russakovsky et al. 2015). We have also evaluated the performance of all comparison methods on the testing set of CIFAR-10 and CIFAR100.
Dataset Splits Yes We evaluate the performance of the comparison methods on Image Net-1K dataset which contains 1000 images sampled from the Image Net-1K validation set (Russakovsky et al. 2015).
Hardware Specification Yes All experiments were performed on a computer with a NVIDIA RTX 3090 GPU with 24 GB memory.
Software Dependencies No No specific version numbers for software dependencies (e.g., Python 3.x, PyTorch 1.x) were found in the paper, only mentions of models like VGG16.
Experiment Setup Yes The optimizer for optimizing the learning objective of Adv INN in (1) is set to Adam (Kingma and Ba 2014) optimizer with initial learning rate 1e 4 which is decayed every 200 iterations with decay rate 0.9 and is lower bounded by 1e 5. We empirically set the regularization parameters λadv = 3, wll = 2, wlh,hl,hh = 1 and λperp = 0.001. All methods use ϵ = 8/255 as the adversarial budget.