Enhancing the Antidote: Improved Pointwise Certifications against Poisoning Attacks

Authors: Shijie Liu, Andrew C. Cullen, Paul Montague, Sarah M. Erfani, Benjamin I. P. Rubinstein

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our defence method achieves more than double the number of poisoned examples compared to existing certified approaches as demonstrated by experiments on MNIST, Fashion-MNIST and CIFAR-10. To verify the effectiveness of our proposed pointwise-certified defence, we conducted experiments across MNIST, Fashion-MNIST, and CIFAR-10 for varying levels of added noise σ.
Researcher Affiliation Collaboration 1School of Computing and Information Systems, University of Melbourne, Melbourne, Australia 2Defence Science and Technology Group, Adelaide, Australia
Pseudocode Yes Algorithm 1: Certifiably Robust Differentially Private Defence Algorithm
Open Source Code No The paper does not contain an explicit statement about releasing the source code or a link to a code repository for the methodology described.
Open Datasets Yes To verify the effectiveness of our proposed pointwise-certified defence, we conducted experiments across MNIST, Fashion-MNIST, and CIFAR-10 for varying levels of added noise σ.
Dataset Splits No The paper specifies training datasets (MNIST, Fashion-MNIST, CIFAR-10) and mentions a 'testing dataset De' but does not explicitly detail the training/test/validation split percentages or methodology for a validation set.
Hardware Specification Yes All experiments were conducted in Pytorch using a single NVIDIA RTX 2080 Ti GPU with 11 GB of GPU RAM.
Software Dependencies No The paper mentions 'Pytorch' but does not specify a version number or list other software dependencies with their versions.
Experiment Setup Yes Across all experiments adjust the sample ratio q to have a batch size of 128, with training conducted using ADAM with a learning rate of 0.01 optimising the Cross-Entropy loss. The clip size C is fine-tuned for each experiment (around 1.0 on MNIST, 25.0 on CIFAR-10). In each case, uncertainties were estimated for a confidence interval suitable for η = 0.001.