Machine Unlearning for Image-to-Image Generative Models
Authors: Guihong Li, Hsiang Hsu, Chun-Fu Chen, Radu Marculescu
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical studies on two large-scale datasets, Image Net1K and Places-365, further show that our algorithm does not rely on the availability of the retain samples, which further complies with data retention policy. |
| Researcher Affiliation | Collaboration | Guihong Li1 , Hsiang Hsu2, Chun-Fu (Richard) Chen2, Radu Marculescu1 1The University of Texas at Austin, USA 2Global Technology Applied Research, JPMorgan Chase, USA |
| Pseudocode | Yes | We provide the details of our unlearning algorithm and corresponding pseudo code in Appendix C.4. |
| Open Source Code | Yes | Our code is available at https://github.com/jpmorganchase/l2lgenerator-unlearning. |
| Open Datasets | Yes | All the datasets used in this paper are open dataset and are available to the public. Besides, our codes are primarily based on Py Torch (Paszke et al., 2019). We use several open source code base and model checkpoints to build our own approach (see Appendix C.1). Our approach can be implemented by obtaining the outputs of target model s encoders and the original model s encoders and then computing the L2-loss between them. We provide more implementation details in Appendix C. |
| Dataset Splits | Yes | We randomly select 100 classes out of 1000 classes as the retain set and another 100 classes as the forget set. We select 100 images per class from the training set as the training set for unlearning. By using the approach defined in Eq. (10), we obtain the target model. We then evaluate the obtained model on both forget set and retain set, with 50 images per class from the Image Net-validation set; hence, we have 5,000 validation images for both the retains set and forget set. |
| Hardware Specification | Yes | Overall, it takes 1.5 hours on a 8 NVIDIA A10G server. Overall, it takes one hour on a 4 NVIDIA A10G server. |
| Software Dependencies | No | The paper mentions "Py Torch (Paszke et al., 2019)" and various generative models it's based on, but does not provide specific version numbers for PyTorch or other software dependencies. |
| Experiment Setup | Yes | We set the learning rate as 10 5 with no weight decay. We use the Adam as the optimizer and conduct the unlearning for 3 epochs. We set the input resolution as 256 and set the batch size as 8 per GPU. Overall, it takes 1.5 hours on a 8 NVIDIA A10G server. We set α = 0.25 (cf. Eq. (10)). |