Model-Targeted Poisoning Attacks with Provable Convergence

Authors: Fnu Suya, Saeed Mahloujifar, Anshuman Suri, David Evans, Yuan Tian

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

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, it either exceeds or matches state-of-the-art attacks in terms of attack success rate and distance to the target model.
Researcher Affiliation Academia 1University of Virginia 2Princeton University. Correspondence to: Fnu Suya <suya@virginia.edu>, Saeed Mahloujifar <sfar@princeton.edu>.
Pseudocode Yes Algorithm 1 Model Targeted Poisoning
Open Source Code Yes All of our evaluation code is available at: https://github.com/suyeecav/model-targeted-poisoning.
Open Datasets Yes For the subpopulation attack experiments, we use the Adult dataset (Dua & Graff, 2017), which was used for evaluation by (Jagielski et al., 2019). [...] For the indiscriminate setting, we use the Dogfish (Koh & Liang, 2017) and MNIST 1 7 datasets (Le Cun, 1998)1.
Dataset Splits No The paper provides training and test set counts but does not explicitly mention or provide details for a separate validation dataset split.
Hardware Specification No The paper states, 'all of our experiments can be run on a typical laptop,' which is too vague to be considered a specific hardware specification.
Software Dependencies No The paper mentions using hinge loss for SVM and logistic loss for LR, and the Adam optimizer, but it does not specify version numbers for any software, libraries, or frameworks used in the experiments.
Experiment Setup Yes For SVM model, we set ϵ as 0.01 on Adult, 0.1 on MNIST 1 7 and 2.0 on Dogfish dataset. For LR model, we set ϵ as 0.05 on Adult, 0.1 on MNIST 1 7 and 1.0 on Dogfish. [...] We use Adam optimizer (Kingma & Ba, 2014) with random restarts to solve this maximization problem approximately.