DifAttack: Query-Efficient Black-Box Adversarial Attack via Disentangled Feature Space

Authors: Jun Liu, Jiantao Zhou, Jiandian Zeng, Jinyu Tian

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results demonstrate that our method achieves significant improvements in ASR and query efficiency simultaneously, especially in the targeted attack and open-set scenarios. Experiments In this section, we compare our Dif Attack (Dif) with SOTA score-based black-box attack methods in both the close-set and open-set scenarios. Ablation experiments on the DF module and τ in Eq.(9) are also conducted.
Researcher Affiliation Academia 1 State Key Laboratory of Internet of Things for Smart City, Department of Computer and Information Science, University of Macau 2 Institute of Artificial Intelligence and Future Networks, Beijing Normal University 3 School of Computer Science and Engineering, Macau University of Science and Technology
Pseudocode Yes Algorithm 1: The proposed Dif Attack method. Input: target classifier F; clean image x; max query budget Q > 0; distortion budget ϵ; ground-truth label y; learning rate η; variance σ; v, k in Eq.(8); sample scale τ. Output: Adversarial image x ....
Open Source Code Yes The code is available at https://github.com/csjunjun/Dif Attack.git.
Open Datasets Yes Datasets. We mainly conduct experiments on the large-scale Image Net-1k, small-scale Cifar-10 and Cifar-100 datasets.
Dataset Splits No We randomly select a target class and 1,000 images, excluding those belonging to the selected class, as the test set. The paper does not specify the training and validation splits.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, memory) used for the experiments.
Software Dependencies No The paper mentions "torchvision2" and links to `https://pytorch.org/vision/stable/index.html` but does not specify exact version numbers for PyTorch or other libraries.
Experiment Setup Yes Parameters. Following many previous works, we also set the maximum perturbation on Image Net to 12/255, and on CIFAR-10 as well as CIFAR-100 to 8/255. The maximum query number for these three datasets is set to 10,000. For the real-world Imagga API, the maximum query number is limited to 500 due to their query limit. In our Dif Attack, we set λ = 1, σ = 0.1, η = 0.01, k = 5 or 0 for Eq.(3) and (9). Additionally, in targeted attacks, we generally set τ to the optimal value of most classifiers, which is 12. In untargeted attacks, τ is usually set to 8.