Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models

Machine unlearning empowers individuals with the `right to be forgotten' by removing their private or sensitive information encoded in machine learning models. However, it remains uncertain whether MU can be effectively applied to Multimodal Large Language Models (MLLMs), particularly in scenarios of forgetting the leaked visual data of concepts. To overcome the challenge, we propose an efficient method, Single Image Unlearning (SIU), to unlearn the visual recognition of a concept by fine-tuning a single associated image for few steps. SIU consists of two key aspects: (i) Constructing Multifaceted fine-tuning data. We introduce four targets, based on which we construct fine-tuning data for the concepts to be forgotten; (ii) Jointly training loss. To synchronously forget the visual recognition of concepts and preserve the utility of MLLMs, we fine-tune MLLMs through a novel Dual Masked KL-divergence Loss combined with Cross Entropy loss. Alongside our method, we establish MMUBench, a new benchmark for MU in MLLMs and introduce a collection of metrics for its evaluation. 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. To the best of our knowledge, we are the first to explore MU in MLLMs. We will release the code and benchmark in the near future.

Paper

Similar papers

Peer review

Reviewer j8ff8/10 · confidence 3/52024-07-07

Summary

The authors propose Single Image Unlearning (SIU), to unlearn the visual recognition of a concept in Multimodal Large Language Models (MLLMs). They also introduce MMUBench, a new benchmark for machine unlearning in MLLMs.

Strengths

The paper introduces a new approach to machine unlearning in MLLMs. The concept of Single Image Unlearning (SIU) is novel, and the additional introduction of MMUBench is essential to quantify methods and push research further. The paper presents a well-structured and technically sound methodology. The proposed SIU method, incorporating Multifaceted Fine-tuning Data and Dual Masked KL-divergence Loss, is well explained and demonstrated in experiments. Especially the good performance when unlearning multiple concepts at once as a batch is significant for practice and a significant step forward from previous methods. Overall, the topic is of high relevance, the method seems practical for real-life usage, and MMUBench is a valuable resource for future works.

Weaknesses

The main limitation of the paper is (as acknowledged by the authors) that they only use a single LLM type (LLAVA).

Questions

How do the benchmarked methods compare in terms of compute time and how do they scale?

Rating

8

Confidence

3

Soundness

3

Presentation

3

Contribution

4

Limitations

Method limitations are given but as far as I see there is no discussion of societal impact.

Reviewer DWvJ5/10 · confidence 3/52024-07-13

Summary

In this paper, the authors proposed a single image unlearning (SIU) approach, which aims at unlearning the concepts recognized from a single training image. A Dual Masked KL-divergence (DMK) loss was introduced to be jointly trained with the cross-entropy loss to mitigate the degradation of MLLMs. Besides, to evaluate the machine unlearning performance, a MMUBench was created, over which the proposed SIU demonstrated effectiveness and robustness.

Strengths

- It is an interesting and practical topic to explore visual recognition of concepts to be forgotten during the model training. - A new benchmark MMUBench was establish for evaluating the machine unlearning performance in multimodal LLMs.

Weaknesses

- It is not quite clear whether the incorrect answer is caused by unlearning or hallucinations. - It would be interesting to see whether the SIU keeps achieving good performance when using different multimodal LLMs. The performance after unlearning can be compared across different MLLMs, which will not hurt the comparison before/after the unlearning. - The social impacts are expected to be discussed in the paper/appendix.

Questions

- When aligning with unseen concepts, the authors mentioned “We assume though an incorrect response might be a hallucination, it actually achieves the purpose of unlearning”. How to ensure that such incorrect answers will not impact the model performance by generating hallucinated answers? In the example illustrated in Figure 1, after unlearning, the MLLM gave the incorrect answer of Jacob Campbell. Wouldn’t it be better to say something like “The central figure in this photograph private/sensitive” instead of giving the incorrect answer? - In Equation 1, wouldn’t the two components, i.e., for the set of data to forget and not to forget, have opposite signs? More specifically, for the first component, why it is minimizing the negative of log likelihood instead of positive? - How long it takes to train the model over four A100 40G GPUs?

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

Please refer to Weakness and Question, which are the aspects suggested to be further improved.

Reviewer tqoU5/10 · confidence 3/52024-07-15

Summary

This paper proposes an algorithm for the unlearning of the visual recognition of concepts from single images in Multimodal Large Language Models (MLLMs) and introduces a benchmark dataset for evaluation. To achieve single image unlearning in MLLMs, the following process is conducted. First, based on the given image, Multifaceted Fine-tuning data is generated for four purposes: Aligning with Unseen Concepts, Assigning New Visual Descriptions, Decoupling Factual Knowledge, and Preserving Non-targeted Knowledge. Second, to ensure effective unlearning using the generated fine-tuning data, a Dual Masked KL-Divergence Loss, which considers token-level and vocabulary-level masking, is proposed. Additionally, the proposed MMUBench is designed as a dataset to evaluate unlearning the visual recognition of concepts from a given single image. Several metrics such as Efficacy and Generality are suggested to assess the success of unlearning. Finally, experimental results using both the proposed algorithm and existing baselines on MMUBench demonstrate that the proposed algorithm achieves superior results.

Strengths

1. As far as I know, this paper is the first to define the problem of unlearning in MLLMs and propose an algorithm and benchmark dataset for evaluation. 2. The algorithm proposed to achieve the unlearning goal set in Line 120 is well-designed with appropriate motivation to ensure successful unlearning. 3. Additionally, the proposed MMUBench offers a well-constructed dataset and various evaluation metrics to assess the achievement of the unlearning goal. 4. Finally, experiments using MMUBench have experimentally validated that the proposed algorithm achieves superior performance in MLLMs compared to existing algorithms.

Weaknesses

The weaknesses of this paper are as follows: 1. The structure of the paper makes it difficult to understand the entire content smoothly. Many crucial details are located in the Appendix, which hampers the reading flow. For example: 1.1) In Section 4.1, the authors refer to Figure 7 in the Appendix for explanation, but this information is also presented in the middle image of Figure 1. It would be better if the authors used Figure 1 to explain as much as possible. 1.2) Many parts that should be in the main text are in the Appendix. For instance, the motivation for DMK in Lines 164-165 should be mentioned in the main text. Additionally, a brief introduction to the MMUBench dataset in Lines 183-184 should be included in the main text. Even if it means reducing content in other parts, including these details in the main text would significantly improve the readability of the paper. 2. Are all the experimental results in the main text reported as the average results of multiple random seeds? It is necessary to conduct experiments with at least five seeds and report the standard deviation. 3. Lastly, what is the reason or motivation for targeting the unlearning of the visual recognition of concepts rather than factual knowledge in MLLMs? What are examples where the proposed unlearning goal is needed in the real world? The authors should include this information in the main text. Personally, although I am not an expert in MLLMs, I believe that unlearning factual knowledge might be a more necessary form of unlearning for MLLMs in the real world.

Questions

Please refer to the Weakness section. I positively evaluate the paper for defining the unlearning goal in MLLMs and proposing the corresponding algorithm and benchmark dataset. However, the paper's structure needs improvement, there is a concern about whether the experimental results consider multiple seeds, and there is a lack of explanation regarding the necessity of the proposed unlearning goal in the real world. These issues make it difficult to give a higher evaluation. If the authors address these concerns appropriately, I would be happy to raise the score.

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

2

Limitations

There is no potential negative societal impact of this paper.

Reviewer j8ff2024-08-07

Thank you for the clarification. No further questions from my side.

Reviewer tqoU2024-08-10

I would like to thank the authors for providing their response. The concerns I raised have mostly been addressed through the author response. However, I believe the paper still needs to be revised to improve readability, and the motivation discussed in response to Q3 should be integrated into the manuscript. Accordingly, I have increased my score from 4 to 5.

Authorsrebuttal2024-08-10

Thank you for reading our rebuttal and for updating the review!

Reviewer DWvJ2024-08-12

Thank you for your detailed clarification. My concerns have mostly been addressed through the author response. I have no further questions at this time and tend to keep my current score.

Authorsrebuttal2024-08-12

Thank you for your thoughtful review and for acknowledging our clarifications. We are committed to addressing all concerns in the final version. Reviewer tqoU has kindly adjusted their score, and Reviewer j8ff has expressed strong support for our work. We have made considerable efforts to address your concerns regarding soundness, particularly through our detailed experimental results, the clarification of our formulas, and the explanation of whether the outputs are due to hallucination. Additionally, our work on defining appropriate target outputs in the unlearning domain and expanding the discussion on societal impacts strengthens the contribution of our study. If you are happy with our response to your comments related to soundness and contribution, we would be deeply grateful if you could consider raising the score as well. Your evaluation is very important to us, and we truly appreciate your consideration. Thank you again for your valuable feedback.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC