TtBA: Two-third Bridge Approach for Decision-Based Adversarial Attack
Authors: Feiyang Wang, Xingquan Zuo, Hai Huang, Gang Chen
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on MNIST, FASHION-MNIST, CIFAR10, CIFAR100, and Image Net datasets demonstrate the strong performance and scalability of our approach. Compared to state-of-the-art non-targeted and targeted attack methods, Tt BA consistently delivers superior performance across most experimented datasets and deep learning models. Code is available at https://github.com/BUPTAIOC/Tt BA. |
| Researcher Affiliation | Academia | 1 School of Computer Science, Beijing University of Posts and Telecommunications, Beijing, China 2 Key Laboratory of Trustworthy Distributed Computing and Services, Ministry of Education, Beijing, China 3School of Engineering and Computer Science, Victoria University of Wellington, Wellington, New Zealand. Correspondence to: Xingquan Zuo <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Two-third Bridge Attack Algorithm 2 Decision Boundary Binary Search Algorithm 3 Get Normal Vector Algorithm 4 Get Initial Perturbation Direction |
| Open Source Code | Yes | Code is available at https://github.com/BUPTAIOC/Tt BA. |
| Open Datasets | Yes | Experimental results on MNIST, FASHION-MNIST, CIFAR10, CIFAR100, and Image Net datasets demonstrate the strong performance and scalability of our approach. 1. MNIST-CNN: The MNIST dataset (Le Cun et al., 1998)... 2. FASHION-MNIST-CNN: The FASHION MNIST dataset (Xiao et al., 2017)... 3. CIFAR10-CNN: The CIFAR10 dataset (Krizhevsky et al., 2009)... 4. CIFAR100-Vi T: The CIFAR100 dataset (Krizhevsky et al., 2009)... 5. The Image Net dataset (Deng et al., 2009)... |
| Dataset Splits | No | For each model, we randomly select 1000 images from the test dataset. For each dataset, we randomly select 500 images to conduct targeted and non-targeted attack experiments, adhering to a query budget limit of 10,000 queries. The paper mentions using images from the 'test dataset' and randomly selecting a subset, but it does not specify the train/validation/test split for the full datasets or cite a predefined split. |
| Hardware Specification | Yes | Experiment hardware configuration. Experiments are conducted using an Intel Xeon Gold 6330 CPU and NVIDIA Ge Force RTX 4090 GPU, running Py Torch 2.3.0, Torchvision 0.18.0, and Python 3.11.5. |
| Software Dependencies | Yes | Experiments are conducted using an Intel Xeon Gold 6330 CPU and NVIDIA Ge Force RTX 4090 GPU, running Py Torch 2.3.0, Torchvision 0.18.0, and Python 3.11.5. |
| Experiment Setup | Yes | Hyperparameter settings. We adopt the recommended hyperparameter settings in (Reza et al., 2023) for searching decision boundaries and normal vectors. Specifically, for all four comparative algorithms and Tt BA, the decision boundary search tolerance τ = 0.0001. The dimension reduction factor is set to s = 4 for the Image Net dataset, and s = 1 for all other datasets. In Tt BA, search tolerance δ = 0.001 for determining ki bridge. According to the parameter sensitive analysis in Appendix G, the thresholds ˇk = 0.1 and ˆk = 0.2, and the bridge coefficients ˆb = 0.9 and ˇb = 2/3. |