Improved Localized Machine Unlearning Through the Lens of Memorization
Authors: Reihaneh Torkzadehmahani, Reza Nasirigerdeh, Georgios Kaissis, Daniel Rueckert, Gintare Karolina Dziugaite, Eleni Triantafillou
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments, we find that our localization strategy outperforms prior strategies in terms of metrics of interest for unlearning and test accuracy, and pairs well with various unlearning algorithms. Our experiments on different datasets, forget sets, and metrics reveal that DEL outperforms prior work in producing better trade-offs between unlearning performance and accuracy. |
| Researcher Affiliation | Collaboration | Reihaneh Torkzadehmahani EMAIL Technical University of Munich Reza Nasirigerdeh EMAIL Technical University of Munich Georgios Kaissis EMAIL Google Deep Mind Technical University of Munich Daniel Rueckert EMAIL Technical University of Munich Imperial College London Gintare Karolina Dziugaite EMAIL Google Deep Mind Mila Eleni Triantafillou EMAIL Google Deep Mind |
| Pseudocode | Yes | A.2 Pseudocode Algorithm 1: Our localization strategy |
| Open Source Code | Yes | The code is available at: https://github.com/reihaneh-torkzadehmahani/DEL-Unlearning/ |
| Open Datasets | Yes | Datasets The CIFAR-10 dataset (Krizhevsky et al., 2009) consists of 50, 000 train and 10, 000 test images of shape 32 32 from 10 classes. The SVHN dataset (Netzer et al., 2011) includes 73, 257 train and 26, 032 test samples. The Image Net-100 (Hugging Face version) dataset is a subset of Image Net (Deng et al., 2009), containing 126, 689 train and 5, 000 test samples from 100 classes, randomly selected from the original Image Net classes. |
| Dataset Splits | Yes | The CIFAR-10 dataset (Krizhevsky et al., 2009) consists of 50, 000 train and 10, 000 test images of shape 32 32 from 10 classes. The SVHN dataset (Netzer et al., 2011) includes 73, 257 train and 26, 032 test samples. The Image Net-100 (Hugging Face version) dataset is a subset of Image Net (Deng et al., 2009), containing 126, 689 train and 5, 000 test samples from 100 classes, randomly selected from the original Image Net classes. |
| Hardware Specification | No | No specific hardware details (like GPU/CPU models or memory) are provided in the paper. The paper only refers to training models without specifying the underlying hardware. |
| Software Dependencies | No | Models We capitalize on the original implementation of Res Net-18 and Res Net-50 (He et al., 2016) from Py Torch and the implementation of Vision Transformer (Vi T) (Dosovitskiy, 2020) from Wang (2021). While PyTorch is mentioned, no version number is provided for it or any other software library. |
| Experiment Setup | Yes | Training For the original (pretrained) models, we train Res Net-18 on CIFAR-10 and Vi T on SVHN (i.e. on the training set of the datasets) for 50 epochs using the SGD optimizer with momentum of 0.9, cross-entropy loss function, and batch size of 128. The base learning rate values are 0.1 and 0.05 for CIFAR-10 / Res Net-18 and SVHN / Vi T, respectively, which are gradually decayed by a factor of 0.01 using the Cosine Annealing scheduler. For the oracle model (gold standard), we train the model from scratch only on the retain set, following the same procedure employed for the pretrained model, except the number of epochs, which we set to 20, and the learning rate, which is half of that in the original training. In the tables below, we provide the hyper-parameter values for the approximate unlearning algorithms. We repeat each experiment three times and report the average values along with 95% confidence interval. |