MultiDelete for Multimodal Machine Unlearning

Authors:

• Jiali Cheng (jiali_cheng@uml.edu)
• Hadi Amiri (hadi_amiri@uml.edu)

MultiDelete Paper: ECCV 2024, Preprint

Overview

We propose MultiDelete, the first machine unlearning method that targets unlearning multimodal data and models (MLLM). It formulates multimodal unlearning as 1) Modality Decoupling, 2) Multimodal Knowledge Retention, 3) Unimodal Knowledge Retention.

How to run

1. Step 1. Train original model

bash bash/ori.sh

2. Step 2. Unlearn

python bash/run.py

Citation

If you find MultiDelete useful for your research, please consider citing this paper:

@inproceedings{cheng2024multidelete,
author="Cheng, Jiali
and Amiri, Hadi",
title="MultiDelete for Multimodal Machine Unlearning",
booktitle="Computer Vision -- ECCV 2024",
year="2025",
publisher="Springer Nature Switzerland",
isbn="978-3-031-72940-9"
}