Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models
Authors: Jiaqi Li, Qianshan Wei, Chuanyi Zhang, Guilin Qi, Miaozeng Du, Yongrui Chen, Sheng Bi, Fan Liu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on MMUBench show that SIU completely surpasses the performance of existing methods. Furthermore, we surprisingly find that SIU can avoid invasive membership inference attacks and jailbreak attacks. |
| Researcher Affiliation | Academia | 1 School of Cyber Science and Engineering, Southeast University, Nanjing, China 2 College of Artificial Intelligence and Automation, Hohai University, Nanjing, China 3 Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China 4School of Computer Science and Engineering, Southeast University, Nanjing, China jqli@seu.edu.cn, 213223283@seu.edu.cn, 20231104@hhu.edu.cn, gqi@seu.edu.cn miaozengdu@seu.edu.cn,yrchen@seu.edu.cn,shengbi@seu.edu.cn,fanliu@hhu.edu.cn |
| Pseudocode | No | The paper does not include a figure, block, or section explicitly labeled "Pseudocode", "Algorithm", or "Algorithm X". |
| Open Source Code | Yes | We include the code and data in our supplemental material. |
| Open Datasets | Yes | We establish MMUBench, a comprehensive benchmark designed to assess MU within MLLMs. This benchmark includes a curated dataset with a minimum of 50 images for each of 20 concepts. One image per concept is designated for the forgetting training set, with the remainder serving to assess generality. ... The construction of dataset is detailed in Appendix C.1. ... Our first step was to sample a diverse set of 300 concepts from the MIKE dataset [22]. |
| Dataset Splits | No | The paper specifies training and testing subsets but does not explicitly mention a separate validation dataset split. |
| Hardware Specification | Yes | We use four A100 40G GPUs to train the model. |
| Software Dependencies | No | The optimizer is Adam and the learning rate is 3e-4. Lora [16] is employed to fine-tune LLAVA with batch size 4. The paper mentions LLAVA (7B and 13B) as models used, but does not provide specific version numbers for software dependencies like PyTorch, CUDA, or other libraries beyond general mentions of optimizer and Lora. |
| Experiment Setup | Yes | The optimizer is Adam and the learning rate is 3e-4. Lora [16] is employed to fine-tune LLAVA with batch size 4. The training step is set to 6. We use four A100 40G GPUs to train the model. α and β are 0.9 and 0.75 respectively. |