Uncovering the Connections Between Adversarial Transferability and Knowledge Transferability

Authors: Kaizhao Liang, Jacky Y Zhang, Boxin Wang, Zhuolin Yang, Sanmi Koyejo, Bo Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments for different scenarios on diverse datasets, showing a positive correlation between adversarial transferability and knowledge transferability.
Researcher Affiliation Academia 1Department of Computer Science, the University of Illinois at Urbana-Champaign, Urbana, USA.
Pseudocode No The paper describes methods and theoretical analysis but does not include any pseudocode blocks or algorithms in a structured format.
Open Source Code Yes All code and data are available here.
Open Datasets Yes All code and data are available here. We first train 5 different architectures (Alex Net, Fully connected network, Le Net, Res Net18, Res Net50) on cifar10 (Krizhevsky et al., 2009). Then we perform transfer learning to STL10 (Coates et al., 2011).
Dataset Splits No The paper does not explicitly state specific training/validation/test splits (e.g., percentages or sample counts) or mention cross-validation setups.
Hardware Specification No The paper does not provide specific details about the hardware used for the experiments, such as GPU models, CPU types, or memory.
Software Dependencies No The paper mentions training models and using certain attack methods (PGD, T3, MI) but does not specify any software or library versions (e.g., PyTorch 1.x, Python 3.x).
Experiment Setup Yes Models Both the source model f S and target model f T are one-hidden-layer neural networks with sigmoid activation. ... we perturb the model weights of the clean source model as W := W + t V ... We use the standard ℓ2 loss as the adversarial loss function. ... we generate adversarial examples with PGD ... generate adversarial examples by the state-of-the-art whitebox attack algorithm T3 ... conduct ablation studies ... on two additional attack methods, MI (Tram er et al., 2017a), PGD-L2 and two additional ϵ with PGD, 2/225, 4/255.