On Success and Simplicity: A Second Look at Transferable Targeted Attacks

Authors: Zhengyu Zhao, Zhuoran Liu, Martha Larson

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we provide experimental evidence to show the general effectiveness of simple transferable attacks. Firstly, in Section 4.1, we evaluate the simple transferable attacks in a variety of transfer scenarios, including single-model transfer, ensemble transfer (easy and challenging scenarios), a worse-case scenario with low-ranked target classes, and a real-world attack on the Google Cloud Vision API.
Researcher Affiliation Academia Zhengyu Zhao, Zhuoran Liu, Martha Larson Radboud University {z.zhao,z.liu,m.larson}@cs.ru.nl
Pseudocode No The paper describes methods using mathematical formulations and descriptive text, but it does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/Zhengyu Zhao/Targeted-Tansfer.
Open Datasets Yes we used the 1000 images from the development set of the Image Net-Compatible Dataset1, which was introduced along with the NIPS 2017 Competition on Adversarial Attacks and Defenses. 1https://github.com/cleverhans-lab/cleverhans/tree/master/cleverhans_v3. 1.0/examples/nips17_adversarial_competition/dataset.
Dataset Splits No The paper mentions using the '1000 images from the development set of the Image Net-Compatible Dataset', but it does not specify explicit training, validation, or test splits for these images within their own experimental setup. The experiments are conducted on pre-trained models.
Hardware Specification Yes Our experiments were run on an NVIDIA Tesla P100 GPU with 12GB of memory.
Software Dependencies No The paper does not provide specific version numbers for software dependencies such as programming languages, libraries, or frameworks used for implementation (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes All attacks used TI, MI, and DI with optimal hyperparameters provided in their original work. Specifically, W 1 = 5 was used for TI as suggested by [12]. If not mentioned specifically, all attacks were run with 300 iterations to ensure convergence. When being executed with a batch size of 20, the optimization process took about three seconds per image. A moderate step size of 2 was used for all attacks... Following the common practice, the perturbations were restricted by L norm with ϵ = 16.