Efficient Source-free Unlearning via Energy-Guided Data Synthesis and Discrimination-Aware Multitask Optimization

Authors: Xiuyuan Wang, Chaochao Chen, Weiming Liu, Xinting Liao, Fan Wang, Xiaolin Zheng

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on three benchmark datasets demonstrate that DSDA outperforms existing unlearning methods, validating its effectiveness and efficiency in source-free unlearning.
Researcher Affiliation Academia 1Zhejiang University, China. Correspondence to: Chaochao Chen <EMAIL>.
Pseudocode Yes The pseudocode are shown in Algorithm 1
Open Source Code No The paper does not contain any explicit statements regarding the availability of source code, nor does it provide a link to a code repository.
Open Datasets Yes We conduct experiments on CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009) and Pins Face Recognition (Hereis, 2024) datasets.
Dataset Splits No The paper mentions using CIFAR-10, CIFAR-100, and Pins Face Recognition datasets, which are standard benchmarks. However, it does not explicitly state the specific training/validation/test splits, percentages, or sample counts used for these datasets within the main text.
Hardware Specification Yes All experiments are conducted on two NVIDIA RTX 3090 GPUs and repeated three times with different random seeds.
Software Dependencies No We implement all experiments in Python 3.9 and use the Py Torch library (Paszke et al., 2019).
Experiment Setup Yes Both the original and retrained models are trained from scratch using a multi-step learning rate scheduler, which begins with a learning rate of 0.01, and optimized with the Adam optimizer (Kingma & Ba, 2014). For a fair comparison, the batch sizes of all methods are set to 256 in Res Net18 and 32 in Vi T.