Toward Efficient Data-Free Unlearning

Authors: Chenhao Zhang, Shaofei Shen, Weitong Chen, Miao Xu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that the proposed ISPF effectively tackles the challenge and outperforms existing methods. We evaluate the effectiveness of the ISPF which is composed of two proposed techniques, the Inhibited Synthesis (IS) and Post Filter (PF), on three widely used benchmark datasets, i.e., SVHN (Netzer et al. 2011), CIFAR-10 and CIFAR100 (Krizhevsky, Hinton et al. 2009).
Researcher Affiliation Academia Chenhao Zhang1, Shaofei Shen1, Weitong Chen2, Miao Xu1* 1University of Queensland 2University of Adelaide EMAIL, EMAIL, EMAIL, EMAIL
Pseudocode No The overall algorithm is placed in Appendix B.
Open Source Code Yes Code https://github.com/Child Eden/ISPF
Open Datasets Yes We evaluate the effectiveness of the ISPF which is composed of two proposed techniques, the Inhibited Synthesis (IS) and Post Filter (PF), on three widely used benchmark datasets, i.e., SVHN (Netzer et al. 2011), CIFAR-10 and CIFAR100 (Krizhevsky, Hinton et al. 2009).
Dataset Splits No We test unlearned models on the forgetting test data Dtest f and the retaining test data Dtest r for obtaining accuracies Af and Ar, respectively. Implementation details are in the Appendix C.
Hardware Specification No No specific hardware details (like GPU models or CPU specifications) are provided in the main text. The paper mentions 'Implementation details are in the Appendix C.' but this appendix is not provided.
Software Dependencies No No specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA) are provided in the main text. The paper mentions 'Implementation details are in the Appendix C.' but this appendix is not provided.
Experiment Setup No No specific hyperparameter values or detailed training configurations are provided in the main text. The paper mentions 'Implementation details are in the Appendix C.' but this appendix is not provided.