Adversarial Mixup Unlearning
Authors: Zhuoyi Peng, Yixuan Tang, Yi Yang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive evaluations across benchmark datasets demonstrate that our method significantly outperforms state-of-the-art approaches, offering a robust solution to machine unlearning. This work not only deepens understanding of unlearning mechanisms but also lays the foundation for effective machine unlearning with mixup augmentation. We conduct a series of experiments to validate the unlearning effectiveness of Mix Unlearn. First, Tables 1 and 2 show that our method consistently outperforms existing baselines across a range of configurations, including both label-agnostic and label-aware settings, as well as data-level and class-level unlearning (while achieving comparable results in MNIST data-level unlearning). |
| Researcher Affiliation | Academia | Zhuoyi Peng Yixuan Tang Yi Yang Department of ISOM, The Hong Kong University of Science and Technology EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology using mathematical equations (e.g., Eq. 1, 2, 3, 5, 6, 7) and textual explanations of the framework components (generator, unlearner, contrastive losses) but does not include any distinct, structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Third, empirical evaluations across multiple benchmark datasets demonstrate that our approach surpasses existing unlearning techniques. We have released the code for Mix Unlearn.2 https://github.com/realjeremybot/Mix Unlearn.git |
| Open Datasets | Yes | Following prior works (Bourtoule et al., 2021; Chen et al., 2023; Shen et al., 2024), we conduct experiments on four datasets: CIFAR10 (Krizhevsky et al., 2009), SVHN (Netzer et al., 2011) and MNIST (Le Cun et al., 1998), FASHION-MNIST (Xiao et al., 2017). We conduct experiments on the Image Net dataset (Deng et al., 2009), using the existing Vi T model (Dosovitskiy et al., 2021) as the model of interest. |
| Dataset Splits | Yes | First, in the Class Level Unlearning setup, all data from class 0 is removed. Second, in the Data-Level Unlearning (Basic) setup, we randomly remove 40% of the training data labeled with classes 5 through 9. We utilize the pre-trained weights of vit-b16-224-in21k , which were trained on Image Net-21K. For our experiments, we use the validation set of Image Net-1K, consisting of 50,000 images... This dataset is further randomly split into a training set of 40,000 images (to construct Forgetting and Remaining set) and a testing set of 10,000 images (for validate testing accuracy). Specifically, we randomly remove 10% to 40% of the training data labeled with classes 5 through 9 to train the initial model. |
| Hardware Specification | Yes | We implement Mix Unlearn in Py Torch, and NVIDIA Ge Force RTX 3090 is used for training. SISA utilizes 4 GPU devices to facilitate unlearning, whereas other methods use only a single device. A time cost comparison for class-level unlearning is conducted using an NVIDIA Ge Force RTX 3090. |
| Software Dependencies | No | The paper states: 'We implement Mix Unlearn in Py Torch'. While 'Py Torch' is a software dependency, a specific version number is not provided, nor are other key software components with versions. |
| Experiment Setup | Yes | In all experiments, we use a batch size of 32. A learning rate of 1e-5 is applied for CIFAR-10 and SVHN, while 5e-5 is used for MNIST and FASHION-MNIST. The mix ratio λ is sampled from beta distribution with α from {0.3, 0.5, 0.75, 1, 1.5}. τgen is searched from {0.05, 0.1, 0.5, 1, 5}, τmix is searched from {1, 10, 20, 50} and τreal is searched from {2, 5, 10, 20, 40}. The sharpen temperature T is set to 0.3. ω is tuned from {0.5, 1, 10, 20, 30}. For the MNIST and FASHION-MNIST datasets, we set the unlearning epoch to 20. For the CIFAR-10 and SVHN datasets, we set the unlearning epoch to 40. |