Hard to Forget: Poisoning Attacks on Certified Machine Unlearning

Authors: Neil G. Marchant, Benjamin I. P. Rubinstein, Scott Alfeld7691-7700

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper we demonstrate how an attacker can exploit this oversight, highlighting a novel attack surface introduced by machine unlearning. We consider an attacker aiming to increase the computational cost of data removal. We derive and empirically investigate a poisoning attack on certified machine unlearning where strategically designed training data triggers complete retraining when removed. and 4 Experiments We investigate the impact of our attack on the computational cost of unlearning in a variety of simulated settings.
Researcher Affiliation Academia Neil G. Marchant,1 Benjamin I. P. Rubinstein,1 Scott Alfeld2 1 School of Computing and Information Systems, University of Melbourne, Parkville, Australia 2 Department of Computer Science, Amherst College, Amherst, MA, USA nmarchant@unimelb.edu.au, brubinstein@unimelb.edu.au, salfeld@amherst.edu
Pseudocode Yes Algorithm 1: Learning algorithm for certified removal and Algorithm 2: Unlearning algorithm for certified removal
Open Source Code Yes Code is available at https://github.com/ngmarchant/attackunlearning.
Open Datasets Yes We consider MNIST (Le Cun et al. 1998) and Fashion MNIST (Xiao, Rasul, and Vollgraf 2017), both of which contain d = 28 28 single-channel images from ten classes. We also generate a smaller binary classification dataset from MNIST we call Binary-MNIST which contains classes 3 and 8. Beyond the image domain, we consider human activity recognition (HAR) data (Anguita et al. 2013).
Dataset Splits No The paper mentions 'train/test splits' but does not explicitly specify validation splits or their proportions.
Hardware Specification No The paper mentions 'HPC-GPGPU Facility' but does not provide specific hardware details like GPU/CPU models or memory specifications.
Software Dependencies No The paper does not list specific software dependencies with version numbers, such as Python libraries or frameworks.
Experiment Setup No Further details about the experiments, including hardware, parameter settings poisoning ratios, and the number of trials in each experiment are provided in Appendix B of our extended paper (Marchant, Rubinstein, and Alfeld 2021).