Targeted Unlearning Using Perturbed Sign Gradient Methods With Applications On Medical Images
Authors: George R. Nahass, Zhu Wang, Homa Rashidisabet, Won Hwa Kim, Sasha Hubschman, Jeffrey C. Peterson, Pete Setabutr, Chad A. Purnell, Ann Tran, Darvin Yi, Sathya N. Ravi
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Across benchmark and real-world clinical imaging datasets, our approach outperforms baselines on both forgetting and retention metrics, including scenarios involving imaging devices and anatomical outliers. This work demonstrates the feasibility of unlearning on clinical imaging datasets and proposes it as a tool for model maintenance in scenarios that require removing the influence of specific data points without full model retraining. 3 Experiments |
| Researcher Affiliation | Academia | George R. Nahass EMAIL Department of Biomedical Engineering Department of Ophthalmology University of Illinois Chicago, Zhu Wang Department of Computer Science University of Illinois Chicago, Won Hwa Kim Computer Science and Engineering Pohang University of Science and Technology, South Korea, Sathya N. Ravi EMAIL Department of Computer Science University of Illinois Chicago |
| Pseudocode | Yes | Algorithm 1: Boundary Search via Perturbed Sign Gradient, Algorithm 2: Model Update via Relabeled Forget Set, Algorithm 3: Configurable Outer Loop with Remain Loss and Logit-Based Labeling |
| Open Source Code | No | Code is available here. The actual link is not present in the provided text excerpt, making it ambiguous as per the instructions. |
| Open Datasets | Yes | The benchmark datasets include CIFAR-10 and Fashion MNIST Krizhevsky et al.; Xiao et al. (2017). The CFP-OS dataset is a three-class dataset consisting of 1500 open-source images, curated from IDRID Porwal et al. (2018), ORIGA Zhang et al. (2010), REFUGE Pachade et al. (2020), and G1020 Bajwa et al. (2020). The MRI dataset includes 3000 images from each of four classes (Normal, Glioma, Meningioma, Pituitary), sampled from the dataset of Nickparvar (2021). Full details on the origin of the open-source datasets is provided in Appendix Table 3. |
| Dataset Splits | Yes | An 80/20 train test split is used across all datasets. |
| Hardware Specification | Yes | All experiments are run on across 3 NVIDIA Ge Force RTX 2080 Ti GPUs. |
| Software Dependencies | No | No specific software versions (e.g., Python, PyTorch, TensorFlow versions) are mentioned in the paper. |
| Experiment Setup | Yes | All training is performed using cross-entropy loss and SGD with momentum 0.9, a learning rate of 1e-3, and L2 regularization with coefficient 1e-4. Data augmentation includes random horizontal flips, random rotations up to 15 , color jittering, and random cropping. An 80/20 train test split is used across all datasets. ... For CIFAR-10 and Fashion MNIST, the All-CNN is trained for 25 and 15 epochs, respectively, with a learning rate of 0.01 and batch size 64. All hyperparameters were selected via grid search. ... For all datasets, we conducted a grid search over λ [0, 1e-4, 1e-3, 1e-2, 1e-1] and γ [0, 1e-4, 1e-1, 1] (as in Eq. (3)) and the best pairing were used. |