Cross-Modal Attention Guided Unlearning in Vision-Language Models

arXiv:2510.07567 · cs.CV · Submitted 2025-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Cross-Modal Attention Guided Unlearning in Vision-Language Models".

Paper summary: Tom: So we're looking at this paper called "Cross-Modal Attention Guided Unlearning in Vision-Language Models." It looks like they're tackling the problem of forgetting private data from these big vision and language models, specifically for visual question answering tasks.

Jane: Right, so the core idea here seems to be that instead of retraining the whole massive model, they want a way to selectively forget certain information linked to specific queries by looking at how the visual and text parts interact during processing.

Lu: Exactly. They're proposing this framework called CAGUL, which uses cross-modal attention scores between the visual and text inputs to guide where they encode the unlearning information in the visual tokens that are least important for a particular query.

Meng: Wait, so how does it actually work? I mean, if we don't retrain the main model parameters, what are we actually modifying? Does this just change the input embeddings before they go to the language model?

Tom: It’s not just changing the input. They have these external modules that are trained separately. There's a discriminator module that checks if an image is paired with a private query, and then there's this visual token encoder that transforms the visual tokens based on those attention scores.

Jane: So, to put it plainly, they use the attention map between the picture and the question to decide which parts of the picture are least relevant when you want to forget something. They conjecture that information in those least attended tokens is useful for unlearning private or non-private queries.

Lu: That's what they're doing. They’re using this idea from how VLMs extract details, where they compress visual info into a small set of highly attended tokens, and applying the opposite logic—focusing on the least attended ones for forgetting.

Meng: Okay, so it’s a selection process based on attention scores leading to an encoding step. But what about training? How do we train these external modules without messing up the billions of parameters in the VLM itself?

Tom: They train two main things: the discriminator module and this visual token encoder. The discriminator is trained with a classification objective, and the encoder is trained using specific loss objectives for both forgetting and retaining data separately.

Jane: For forgetting, they use something called Preference Optimization to align the language model's outputs for forget queries with refusal responses instead of the actual correct answer. That’s a clever way to signal what we want to erase.

Lu: And for retaining data, they have an objective that tries to maintain the base model’s predictive utility, which is important because they don't just want to erase everything; they want to keep general knowledge intact.

Meng: I see the trade-off there. So you're trying to find this sweet spot where you forget the private data without destroying the model’s ability to answer normal questions correctly? That sounds like it could be a big deal for production systems.

Paper summary: Tom: Experimental results show that CAGUL performs better than finetuning baselines on forgetting quality and even keeps the retain utility of the finetuned base model. And what's cool is they don't alter any of the pre-trained VLM parameters or incur retraining costs.

Jane: That’s significant because it means you can exercise that right to be forgotten for specific subsets of data without having to spend a lot of time and compute on full finetuning. And they kept the computational efficiency quite high, training far fewer parameters than other methods.

Lu: The ablation study shows that while CAGUL can get good trade-offs between forgetting and retaining metrics without the discriminator, its generalization ability suffers if you remove it. But without this encoder module, forget and retain performance drops significantly.

Meng: So it seems the discriminator is crucial for knowing *when* to apply the token selection mechanism, and the encoder is necessary to actually embed that unlearning signal into those selected tokens. That’s a lot of moving parts to get right at scale.

Tom: This whole CAGUL framework is essentially a way for individual users to exercise the right to be forgotten on subsets of their sensitive queries within these vision-language models. It’s a novel setting for machine unlearning in this domain.

Jane: So, the paper "Cross-Modal Attention Guided Unlearning in Vision-Language Models" proposes using cross-modal attention to guide the encoding of unlearning information into low-importance visual tokens for relevant queries.

Lu: That's the main mechanism they are investigating, focusing on how those attention scores between vision and text dictate which visual tokens get modified for forgetting.

Meng: It’s interesting because it leverages existing attention mechanisms instead of completely rebuilding the model training loop or fine-tuning every layer we have to touch. That makes deployment much more practical.

Tom: The authors show this method works on the FIUBench dataset, using two different open-source vision-language models, and they show it achieves good trade-offs compared to finetuning methods without altering the pre-trained model parameters or retraining costs.

Jane: So what this means for us is that we can potentially make these powerful multimodal models more private and compliant with privacy regulations by using targeted unlearning instead of broad fine-tuning.

Lu: It shifts the focus from massive parameter updates to targeted modifications guided by attention patterns, which is a different direction in how we think about model modification.

Meng: If this scales well on real traffic, it could be a huge win for deploying these models where data privacy is a major concern, because the cost of unlearning seems much lower.

Tom: The conclusion of this work is that CAGUL provides an effective way to handle machine unlearning in vision-language models by using a classifier to identify images from the forget set, an attention-based selection mechanism, and an encoder for embedding information into low-attention visual tokens.

Jane: Essentially, they’ve found a novel setting where users can forget specific subsets of attributes available to the model by leveraging cross-modal attention. It’s a practical approach to privacy in multimodal AI.

Conclusion: Tom: So, we're wrapping up on "Cross-Modal Attention Guided Unlearning in Vision-Language Models." Basically, they showed how you can selectively forget data from these big vision and language models without having to retrain the whole thing from scratch.

Jane: Right, so the title is Cross-Modal Attention Guided Unlearning in Vision-Language Models. It’s about using that attention score between the picture and the text to guide where they put information when you want to forget something.

Lu: They use a discriminator module and a visual token encoder to do this selective forgetting based on how important those tokens are for the specific query you’re asking about.

Meng: I just want to know what this means in practice, Tom. If someone wants to forget their private data, does this mean we can just point at the picture and type the query? Or is there a more complex setup?

Tom: It's more complex than that. You need that discriminator module to first figure out if the image is actually paired with a private query before you even look at those attention scores for token selection.

Jane: And they showed it works on two different vision and language models, meaning this approach isn't tied to just one specific model architecture.

Lu: The key result is that this method achieves a good balance between forgetting the data and keeping the model smart enough to answer regular questions, which is what we call retain utility.

Meng: That balance sounds important because usually when you try to erase something, you either lose too much general knowledge or you forget too much of what's actually private.

Tom: Exactly. They found a sweet spot where it performs better than traditional finetuning methods on the forgetting side while preserving that retained utility.

Jane: So the implication is that for anyone dealing with sensitive data, this offers a new way to exercise the right to be forgotten on specific parts of what a multimodal AI knows.

Lu: It’s about moving away from massive parameter updates toward targeted modifications guided by how the model processes information cross-modally.

Meng: That makes sense for deployment because it doesn't require retraining the entire pre-trained model, which saves a ton of computational power and time at scale.

Tom: It’s definitely more efficient than those full finetuning baselines they compared it against, especially when you look at the training costs.

Jane: And this sets up a lot of interesting work for future research on how to make these models more private without sacrificing their overall usefulness for general tasks.

Karuna Bhaila, Aneesh Komanduri, Minh-Hao Van, Xintao Wu

University of Arkansas

cs.CV

Submitted: 2025-10-08

Updated: 2026-11-14

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: Vision-Language Models (VLMs) have demonstrated immense capabilities in multi-modal understanding and inference tasks such as Visual Question Answering (VQA)…

Key concepts

Cross-Modal Attention Scores
These are the numerical scores calculated between visual input tokens (images) and text input tokens (queries). They measure how closely related a specific part of an image is to a specific word in the prompt. The method uses these scores to decide which visual information is most important for unlearning.
Discriminator Module
This module acts as a classifier that determines whether a given visual input is associated with a query about private data or not. It helps the framework identify which images need to be targeted for the unlearning process, guiding the subsequent token selection.
Visual Token Encoder (Fψ)
This component takes visual tokens deemed important for unlearning and linearly transforms them. By focusing on tokens with low cross-modal attention scores, this encoder embeds the specific information needed to forget data without significantly disrupting the model's overall behavior.

Terminology

Summary

Vision-Language Models (VLMs) have demonstrated immense capabilities in multi-modal understanding and inference tasks such as Visual Question Answering (VQA), which requires models to infer outputs based on visual and textual context simultaneously. p. 0

Where the paper says it

Cross-Modal Attention Guided Unlearning in Vision-Language Models Karuna Bhaila, Aneesh Komanduri, Minh-Hao Van, Xintao Wu Department of Electrical Engineering and Computer Science University of Arkansas {kbhaila,akomandu,haovan,xintaowu}@uark. edu Abstract Vision-Language Models (VLMs) have demonstrated immense capabilities in multi-modal understanding and inference tasks such as Visual Question Answering (VQA), which requires models to infer outputs based on visual and textual context simultaneously.

Our method is motivated by the relationship between visual and text tokens represented as cross-modal attention scores computed between the two modes of input. p. 0
Where the paper says it

Based on this problem definition, we propose a lightweight and efficient method to achieve VLM unlearning. Our method is motivated by the relationship between visual and text tokens represented as cross-modal attention scores computed between the two modes of input.

We propose Cross-Modal Attention Guided Unlearning (CAGUL), a lightweight and efficient VLM unlearning framework that utilizes external modules to encode unlearning information in visual tokens of low importance for relevant queries. p. 0
Where the paper says it

We explore the role of visual tokens for output generation in VLMs using crossmodal attention and utilize it to formulate Cross-Modal Attention Guided Unlearning (CAGUL), a lightweight and efficient VLM unlearning framework. In contrast to computationally expensive model finetuning methods, CAGUL utilizes external modules to encode unlearning information in visual tokens of low importance for relevant queries.

We find that the transformed visual tokens not only prevent leakage but also retain reference model behavior. p. 0
Where the paper says it

In contrast to computationally expensive model finetuning methods, CAGUL utilizes external modules to encode unlearning information in visual tokens of low importance for relevant queries. We find that the transformed visual tokens not only prevent leakage but also retain reference model behavior.

Framework Components

The CAGUL framework is composed of two main components: a discriminator module to detect whether or not a given visual input is paired with a query about private data, and a cross-modal attention-based visual token selection mechanism. Then, for visual inputs predicted to be paired with private data, we use an MLP encoder to linearly transform the visual tokens with the lowest cross-modal attention scores. p. 1

Where the paper says it

First, a discriminator module detects whether or not a given visual input is paired with a query about private data. Then, for visual inputs predicted to be paired with private data, we use an MLP encoder to linearly transform the visual tokens with the lowest cross-modal attention scores.

The final input to the LLM can be represented as Yˆ = LLMθ(Zˆv, Zq) where Zˆv is the set of final visual token embeddings. p. 3
Where the paper says it

We then replace the k image token embeddings in Zv with the transformed representation as follows: Zˆ v,i = Fψ(Zv,i) if i ∈ K Zv,i otherwise (8) where Zˆ v is the set of final visual token embeddings. The final input to the LLM can be represented as follows: Yˆ = LLMθ(Zˆ v, Zq) (9) Unlearning Objective.

Unlearning Objectives

The framework has two trainable modules: the discriminator Cϕ and the visual token encoder Fψ. The discriminator is trained with a classification objective defined as Lbce = EZv log pCϕ(lZv). p. 3

Where the paper says it

Our CAGUL framework has two trainable modules, the discriminator Cϕ and the visual token encoder Fψ. The discriminator is trained with the following classification objective Lbce = EZv log pCϕ (l|Zv) (10) The encoder is trained with specific unlearning objectives defined on the forget and retain sets separately.

The encoder is trained with specific unlearning objectives defined on the forget and retain sets separately. p. 3
Where the paper says it

Our CAGUL framework has two trainable modules, the discriminator Cϕ and the visual token encoder Fψ. The discriminator is trained with the following classification objective Lbce = EZv log pCϕ (l|Zv) (10) The encoder is trained with specific unlearning objectives defined on the forget and retain sets separately.

p. 3
Where the paper says it

Our CAGUL framework has two trainable modules, the discriminator Cϕ and the visual token encoder Fψ. The discriminator is trained with the following classification objective Lbce = EZv log pCϕ (l|Zv) (10) The encoder is trained with specific unlearning objectives defined on the forget and retain sets separately.

Specifically, the objective for retain samples is defined as Lr = E(V,X,Y)∼Dr EY t log pθ(Y tV, X, Y <t). p. 3
Where the paper says it

The discriminator is trained with the following classification objective Lbce = EZv log pCϕ (l|Zv) (10) The encoder is trained with specific unlearning objectives defined on the forget and retain sets separately. Generally, the unlearning objective for retain samples is to maintain the base model’s predictive utility.

For the forget objective in the VQA task, we leverage Preference Optimization (PO) to realize this by aligning the language model’s outputs for forget queries with refusal responses instead of the ground truth. p. 3
Where the paper says it

We leverage Preference Optimization (PO) (Maini et al. 2024) to realize this objective in CAGUL by aligning the language model’s outputs for forget queries with refusal responses instead of the ground truth.

The joint objective for CAGUL is formulated as L = Lbce + Lf + Lr. p. 3
Where the paper says it

We compute the standard causal language modeling loss with the substituted response as Lf = E(V,X,Y )∼Df EY¯ t log pθ(Y¯ t |V, X, Y¯ <t) (12) where Y¯ is a refusal response such as “I cannot answer this question. ” Finally, we formulate the joint objective for CAGUL as L = Lbce + Lf + Lr (13)

Performance and Efficiency

Experimental results show that CAGUL performs better or on par with finetuning-based baselines without altering the pre-trained model parameters or incurring retraining costs. For LLaMA-3. p. 1

Where the paper says it

Formally, we can define the attention weights as a function of the query and key matrices: A = softmax QKT √ d (2) In the context of VLMs, cross-modal attention is implemented as either cross-attention or joint self-attention. For cross-attention-based models like LLaMA-3.

2-11B-Vision-Instruct, CAGUL achieves lower Rouge and EM scores on Df compared to Finetune. p. 6
Where the paper says it

CAGUL Rouge and EM scores on Df demonstrate effective forgetting as noted by the significant drops in these metrics compared to Finetune. The MIA score MinK is also significantly lower than Finetune, which indicates that unlearned models achieved with CAGUL can successfully unlearn information from Df .

Furthermore, CAGUL successfully preserves the retain utility of the finetuned base model. p. 6
Where the paper says it

We report retain metrics on the full retain set Dr and the non-private subset Dnp corresponding to the individuals in Df . We observe that CAGUL successfully preserves the retain utility of the finetuned base model.

Computational efficiency shows that CAGUL trains the fewest number of parameters as the pre-trained VLM parameters are kept frozen and completes training in significantly less time. p. 6
Where the paper says it

The baselines require fewer parameters than Retrain, as we implement LoRA finetuning for these methods as opposed to the full finetuning used for Retrain. Nonetheless, CAGUL trains the fewest number of parameters as the pre-trained VLM parameters are kept frozen in this framework and completes training in significantly less time.

Ablation Insights

The ablation study indicates that CAGUL can achieve good trade-offs between forget and retain metrics without Cϕ, but its generalization ability is negatively affected. Conversely, CAGUL w/o Fψ shows a significant decline in forget and retain performance. p. 7

Where the paper says it

We note that the rouge score for Dr shows significant improvement as the encoder is trained on the full retain set, enabling the model to memorize additional knowledge. Conversely, we observe a significant decline in forget and retain performance for CAGUL w/o Fψ, which shows the importance of Fψ in our framework.

The random selection of visual tokens for perturbation achieves mostly comparable performance for forget metrics but underperforms significantly for retain metrics. p. 7
Where the paper says it

The random selection of visual tokens achieves mostly comparable performance for forget metrics, but underperforms significantly for retain metrics. These results validate the selection of the least attended tokens for perturbation in CAGUL.

Conclusion

In this work, we proposed a novel setting for unlearning in vision-language models where each user can exercise the Right to be Forgotten for any subset of their sensitive queries. We proposed Cross-Modal Attention Guided Unlearning (CAGUL), a framework consisting of a classifier to identify images appearing in the forget set, a cross-modal attention-based visual token selection mechanism, and a visual token encoder to embed unlearning information in visual tokens having low correlation with the query prompt. p. 8

Where the paper says it

8 Conclusion In this work, we proposed a novel setting for unlearning in vision-language models where each user can exercise the Right to be Forgotten for any subset of their sensitive queries. We proposed Cross-Modal Attention Guided Unlearning (CAGUL), a framework consisting of a classifier to identify images appearing in the forget set, a cross-modal attention-based visual token selection mechanism, and a visual token encoder to embed unlearning information in visual tokens having low correlation with the query prompt.

We conducted experiments on the FIUBench dataset using two open-source vision-language models and show that our method achieves desirable trade-offs in performance compared to finetuning-based baseline methods. p. 8
Where the paper says it

We proposed Cross-Modal Attention Guided Unlearning (CAGUL), a framework consisting of a classifier to identify images appearing in the forget set, a cross-modal attention-based visual token selection mechanism, and a visual token encoder to embed unlearning information in visual tokens having low correlation with the query prompt. We conducted experiments on the FIUBench dataset using two open-source vision-language models and show that our method achieves desirable trade-offs in performance compared to finetuning-based baseline methods.

References

Beniaguev, D. 2022. Synthetic Faces High Quality (SFHQ) dataset. p. 9

Where the paper says it

References Beniaguev, D. Synthetic Faces High Quality (SFHQ) dataset.

Bhaila, K.; Van, M.-H.; and Wu, X. 2025. Soft Prompting for Unlearning in Large Language Models. p. 9

Where the paper says it

; and Wu, X. Soft Prompting for Unlearning in Large Language Models.

Bourtoule, L. p. 0

Where the paper says it

In the LLM landscape, research efforts have been made to mitigate privacy leakage in different ways, including machine unlearning. Specifically, machine unlearning addresses concerns regarding leakage of training data as stipulated by the California Consumer Privacy Act (CCPA) and GDPR’s Right to be Forgotten (Cao and Yang 2015; Bourtoule et al.

; Chandrasekaran, V. p. 9
Where the paper says it

Bourtoule, L. ; Chandrasekaran, V.

; Choquette-Choo, C. p. 9
Where the paper says it

; Chandrasekaran, V. ; Choquette-Choo, C.

A.; Jia, H.; Travers, A. p. 9
Where the paper says it

; Choquette-Choo, C. ; Travers, A.

; Zhang, B. p. 0
Where the paper says it

However, training data collected from such sources may contain personally identifiable and/or sensitive information, raising significant privacy concerns, especially with the added complexity of visual signals. For instance, processing visual inputs in VLMs can unintentionally disclose information such as location cues, identity of individuals in the background, etc (Zhang et al.

; Lie, D.; and Papernot, N. p. 9
Where the paper says it

; Zhang, B. ; and Papernot, N.

2021. Machine unlearning. p. 0
Where the paper says it

However, the models may memorize private and/or sensitive information during training and regurgitate it in inference. Recently, machine unlearning has been leveraged to address the leakage of private data in LLMs.

Cao, Y.; and Yang, J. p. 0

Where the paper says it

In the LLM landscape, research efforts have been made to mitigate privacy leakage in different ways, including machine unlearning. Specifically, machine unlearning addresses concerns regarding leakage of training data as stipulated by the California Consumer Privacy Act (CCPA) and GDPR’s Right to be Forgotten (Cao and Yang 2015; Bourtoule et al.

2015. Towards making systems forget with machine unlearning. p. 9
Where the paper says it

; and Yang, J. Towards making systems forget with machine unlearning.

Chakraborty, T. p. 1

Where the paper says it

(2024b) study how image-based manipulations can cause breaches in safety-aligned VLMs. Chakraborty et al.

; Shayegani, E. p. 9
Where the paper says it

Chakraborty, T. ; Shayegani, E.

; Cai, Z.; Abu-Ghazaleh, N. p. 9
Where the paper says it

; Shayegani, E. ; Abu-Ghazaleh, N.

B.; Asif, M. S.; Dong, Y.; Roy-Chowdhury, A. p. 9
Where the paper says it

; Abu-Ghazaleh, N. ; Roy-Chowdhury, A.

; and Song, C. p. 9
Where the paper says it

; Roy-Chowdhury, A. ; and Song, C.

2024. Can Textual Unlearning Solve Cross-Modality Safety Alignment? p. 9
Where the paper says it

; and Song, C. Can Textual Unlearning Solve Cross-Modality Safety Alignment?

Chen, J.; and Yang, D. p. 0

Where the paper says it

In the LLM landscape, research efforts have been made to mitigate privacy leakage in different ways, including machine unlearning. Specifically, machine unlearning addresses concerns regarding leakage of training data as stipulated by the California Consumer Privacy Act (CCPA) and GDPR’s Right to be Forgotten (Cao and Yang 2015; Bourtoule et al.

2023. Unlearn What You Want to Forget: Efficient Unlearning for LLMs. p. 9
Where the paper says it

; and Yang, D. Unlearn What You Want to Forget: Efficient Unlearning for LLMs.

Cheng, J.; and Amiri, H. p. 1

Where the paper says it

VLM unlearning is a relatively new frontier with increased complexity due to the integration of visual signals. Cheng and Amiri (2023) propose a framework for multimodal unlearning in VLMs by decoupling the text and visual modalities.

2023. Multimodal Machine Unlearning. p. 9
Where the paper says it

; and Amiri, H. Multimodal Machine Unlearning.

Ding, C.; Wu, J.; Yuan, Y.; Lu, J.; Zhang, K. p. 0

Where the paper says it

However, training data collected from such sources may contain personally identifiable and/or sensitive information, raising significant privacy concerns, especially with the added complexity of visual signals. For instance, processing visual inputs in VLMs can unintentionally disclose information such as location cues, identity of individuals in the background, etc (Zhang et al.

; Su, A.; Wang, X.; and He, X. 2025. Unified Parameter-Efficient Unlearning for LLMs. p. 9
Where the paper says it

; and He, X. Unified Parameter-Efficient Unlearning for LLMs.

Faraglia, D. p. 4

Where the paper says it

1 Dataset We primarily investigate VLM unlearning on the FIUBench dataset (Ma et al. 2025) composed of 400 unique synthetic images obtained from the SFHQ dataset (Beniaguev 2022), each paired with fictitious biographical information including name, birthdate, address, phone number, occupation, income, health, and criminal records sourced from (Patil 2024; Mendes 2020; Vyas 2017; Faraglia 2024).

2024. Faker.

Fu, C.; Chen, P.; Shen, Y.; Qin, Y.; Zhang, M. p. 0

Where the paper says it

However, training data collected from such sources may contain personally identifiable and/or sensitive information, raising significant privacy concerns, especially with the added complexity of visual signals. For instance, processing visual inputs in VLMs can unintentionally disclose information such as location cues, identity of individuals in the background, etc (Zhang et al.

; Lin, X.; Qiu, Z.; Lin, W.; Yang, J.; Zheng, X. p. 9
Where the paper says it

; Zhang, M. ; Zheng, X.

; Li, K., Sun., and Ji., R. 2023. MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models. p. 9
Where the paper says it

; and Ji, R. MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models.

Gao, L.; Niu, Y.; Tang, T.; Avestimehr, S. p. 9

Where the paper says it

CoRR, abs/2306. ; Avestimehr, S.

; and Annavaram, M. p. 9
Where the paper says it

; Avestimehr, S. ; and Annavaram, M.

2024. Ethos: Rectifying language models in orthogonal parameter space. p. 9
Where the paper says it

; and Annavaram, M. Ethos: Rectifying language models in orthogonal parameter space.

Gong, Y.; Ran, D.; Liu, J.; Wang, C.; Cong, T.; Wang, A.; Duan, S.; and Wang, X. p. 9

Where the paper says it

arXiv preprint arXiv:2403. ; and Wang, X.

2023. FigStep: Jailbreaking Large Vision-language Models via Typographic Visual Prompts. p. 9
Where the paper says it

; and Wang, X. FigStep: Jailbreaking Large Vision-language Models via Typographic Visual Prompts.

Kaduri, O.; Bagon, S. p. 0

Where the paper says it

Our method is motivated by the relationship between visual and text tokens represented as cross-modal attention scores computed between the two modes of input. Kaduri, Bagon, and Dekel (2024) empirically demonstrated that VLMs extract fine-grained details and attributes from visual tokens in a spatially localized manner and compress visual information into a small subset of highly attended tokens.

; and Dekel, T. p. 0
Where the paper says it

Our method is motivated by the relationship between visual and text tokens represented as cross-modal attention scores computed between the two modes of input. Kaduri, Bagon, and Dekel (2024) empirically demonstrated that VLMs extract fine-grained details and attributes from visual tokens in a spatially localized manner and compress visual information into a small subset of highly attended tokens.

2024. What’s in the Image? p. 0
Where the paper says it

Generally, VLMs consist of a vision model that extracts visual features, a cross-modal component that projects the visual features to the LLM embedding space, and a language model that processes projected visual tokens along with input text tokens for output generation. Together, these components enable VLMs to process both image and text modalities and be utilized for a wide range of visio-linguistic tasks.

A Deep-Dive into the Vision of Vision Language Models. p. 9
Where the paper says it

What’s in the Image? A Deep-Dive into the Vision of Vision Language Models.

Li, J.; Wei, Q.; Zhang, C. p. 0

Where the paper says it

However, training data collected from such sources may contain personally identifiable and/or sensitive information, raising significant privacy concerns, especially with the added complexity of visual signals. For instance, processing visual inputs in VLMs can unintentionally disclose information such as location cues, identity of individuals in the background, etc (Zhang et al.

; Qi, G.; Du, M.; Chen, Y.; Bi., and Liu., F. 2024a. Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models. p. 9
Where the paper says it

; and Liu, F. Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models.

Li, N.; Pan, A.; Gopal, A. p. 9

Where the paper says it

, Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024. ; Gopal, A.

; Yue, S.; Berrios, D. p. 9
Where the paper says it

; Gopal, A. ; Berrios, D.

; Gatti, A. p. 9
Where the paper says it

; Berrios, D. ; Gatti, A.

; Li., J. D.; Dombrowski, A. p. 9
Where the paper says it

; Gatti, A. ; Dombrowski, A.

-K., Goel., S.; Phan., L., et al. 2024b. The wmdp benchmark: Measuring and reducing malicious use with unlearning. p. 9
Where the paper says it

; Dombrowski, A. The wmdp benchmark: Measuring and reducing malicious use with unlearning.

Liu, S.; Yao, Y.; Jia, J.; Casper, S. p. 9

Where the paper says it

Association for Computational Linguistics. ; Casper, S.

; Baracaldo, N. p. 9
Where the paper says it

; Casper, S. ; Baracaldo, N.

; Hase, P.; Yao., Y., Liu., C. Y.; Xu., X., Li., H., et al. 2024a. Rethinking machine unlearning for large language models. p. 9
Where the paper says it

; Baracaldo, N. Rethinking machine unlearning for large language models.

Liu, X.; Zhu, Y.; Gu, J.; Lan, Y.; Yang, C.; and Qiao, Y. p. 9

Where the paper says it

arXiv preprint arXiv:2402. ; and Qiao, Y.

2024b. MM-SafetyBench: A Benchmark for Safety Evaluation of Multimodal Large Language Models. p. 9
Where the paper says it

; and Qiao, Y. MM-SafetyBench: A Benchmark for Safety Evaluation of Multimodal Large Language Models.

Liu, Z.; Dou, G.; Jia., M.; Tan, Z.; Zeng, Q.; Yuan, Y.; and Jiang, M. p. 9

Where the paper says it

arXiv:2311. ; and Jiang, M.

2025. Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench. p. 9
Where the paper says it

; and Jiang, M. Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench.

Ma, Y.; Wang, J.; Wang, F.; Ma., S.; Li., J.; Pan., J.; Li., X.; Huang, F. p. 9

Where the paper says it

Association for Computational Linguistics. ; Huang, F.

; Sun, L.; Li., B.; Choi., Y.; Chen., M; and Xiao. p. 9
Where the paper says it

; Huang, F. ; and Xiao, C.

, C. 2025. Benchmarking Vision Language Model Unlearning via Fictitious Facial Identity Dataset. p. 9
Where the paper says it

; and Xiao, C. Benchmarking Vision Language Model Unlearning via Fictitious Facial Identity Dataset.

Maini, P.; Feng, Z.; Schwarzschild, A. p. 9

Where the paper says it

In The Thirteenth International Conference on Learning Representations. ; Schwarzschild, A.

; Lipton, Z. p. 9
Where the paper says it

; Schwarzschild, A. ; Lipton, Z.

C.; and Kolter, J. p. 9
Where the paper says it

; Lipton, Z. ; and Kolter, J.

Z. 2024. TOFU: A task of fictitious unlearning for LLMs. p. 9
Where the paper says it

; and Kolter, J. TOFU: A task of fictitious unlearning for LLMs.

Mendes, B., 2020. NYPD Crime Complaint Data Historic (2006-2019). p. 9

Where the paper says it

First Conference On Language Modeling. NYPD Crime Complaint Data Historic (2006-2019).

Patil, P., 2024. Healthcare Dataset. p. 9

Where the paper says it

NYPD Crime Complaint Data Historic (2006-2019). Healthcare Dataset.

Pawelczyk, M. p. 1

Where the paper says it

2024; Chen and Yang 2023; Zhang et al. Others utilize adaptation techniques specific to LLMs, such as achieving unlearning through in-context learning (Pawelczyk, Neel, and Lakkaraju 2024), manually crafted system prompts (Thaker, Maurya, and Smith 2024), or prompt tuning with unlearning-specific losses to guide generation (Bhaila, Van, and Wu 2025).

; Neel, S.; and Lakkaraju, H. p. 1
Where the paper says it

2024; Chen and Yang 2023; Zhang et al. Others utilize adaptation techniques specific to LLMs, such as achieving unlearning through in-context learning (Pawelczyk, Neel, and Lakkaraju 2024), manually crafted system prompts (Thaker, Maurya, and Smith 2024), or prompt tuning with unlearning-specific losses to guide generation (Bhaila, Van, and Wu 2025).

2024. InContext Unlearning: Language Models as Few-Shot Unlearners. p. 9
Where the paper says it

; and Lakkaraju, H. InContext Unlearning: Language Models as Few-Shot Unlearners.

Samson, L.; Barazani, N. p. 9

Where the paper says it

In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. ; Barazani, N.

; Ghebreab, S. p. 9
Where the paper says it

; Barazani, N. ; Ghebreab, S.

; and Asano, Y. p. 9
Where the paper says it

; Ghebreab, S. ; and Asano, Y.

M. 2024. Privacy-Aware Visual Language Models. p. 9
Where the paper says it

; and Asano, Y. Privacy-Aware Visual Language Models.

Shi, W.; Ajith, A. p. 9

Where the paper says it

arXiv:2405. ; Ajith, A.

; Xia, M.; Huang, Y. p. 9
Where the paper says it

Association for Computational Linguistics. ; Huang, F.

; Liu., D.; Blevins, T. p. 9
Where the paper says it

; Huang, Y. ; Blevins, T.

; Chen., D; and Zettlemoyer, L. p. 9
Where the paper says it

; Blevins, T. ; and Zettlemoyer, L.

2023. Detecting Pretraining Data from Large Language Models. p. 9
Where the paper says it

; and Zettlemoyer, L. Detecting Pretraining Data from Large Language Models.

Thaker, P.; Maurya, Y. p. 1

Where the paper says it

2024; Chen and Yang 2023; Zhang et al. Others utilize adaptation techniques specific to LLMs, such as achieving unlearning through in-context learning (Pawelczyk, Neel, and Lakkaraju 2024), manually crafted system prompts (Thaker, Maurya, and Smith 2024), or prompt tuning with unlearning-specific losses to guide generation (Bhaila, Van, and Wu 2025).

; and Smith, V. p. 1
Where the paper says it

2024; Chen and Yang 2023; Zhang et al. Others utilize adaptation techniques specific to LLMs, such as achieving unlearning through in-context learning (Pawelczyk, Neel, and Lakkaraju 2024), manually crafted system prompts (Thaker, Maurya, and Smith 2024), or prompt tuning with unlearning-specific losses to guide generation (Bhaila, Van, and Wu 2025).

2024. Guardrail baselines for unlearning in LLMs. p. 9
Where the paper says it

; and Smith, V. Guardrail baselines for unlearning in LLMs.

Tomekc¸e, B. p. 1

Where the paper says it

There have also been several studies benchmarking VLM unlearning. Tomekc¸e et al.

; Vero, M.; Staab, R. p. 9
Where the paper says it

Tomekc¸e, B. ; Staab, R.

; and Vechev, M. p. 9
Where the paper says it

; Staab, R. ; and Vechev, M.

2024. Private Attribute Inference from Images with Vision-Language Models. p. 9
Where the paper says it

; and Vechev, M. Pri- ¨ vate Attribute Inference from Images with Vision-Language Models.

Vaswani, A. p. 1

Where the paper says it

We refer to the attention scores between visual and text tokens as cross-modal attention. The attention mechanism (Vaswani et al.

; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L. u., and Polosukhin, I. 2017. Attention is All you Need. p. 0
Where the paper says it

For instance, a user may want to remove references to sensitive information (e. , social security number) but may not necessarily need other general profile information to be removed (e.

Vyas., 2017. USA Housing. p. 9

Where the paper says it

Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL 2025 - Volume 1: Long Papers, Albuquerque, New Mexico, USA, April 29 - May 4, 2025, 4105–4135.

Yao, J.; Chien, E.; Du, M.; Niu, X.; Wang, T.; Cheng., Z.; and Yue., X. 2024. Machine Unlearning of Pre-trained Large Language Models. p. 9

Where the paper says it

; and Liu, F. Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models.

Yao., Y.; Xu., X., and Liu., Y. 2023. Large language model unlearning. p. 9

Where the paper says it

; and Wu, X. Soft Prompting for Unlearning in Large Language Models.

Zhang, R.; Lin, L.; Bai, Y.; and Mei, S. 2024. Negative preference optimization: From catastrophic collapse to effective unlearning. p. 3

Where the paper says it

In CAGUL, we define this objective as the standard causal language modeling loss for for all samples in Dr, including the non-private samples for images appearing in the forget set formulated as Lr = E(V,X,Y )∼Dr EY t log pθ(Y t |V, X, Y <t) (11) In the VQA task, the forget objective can be interpreted as obtaining low utility for samples in the forget set Df . We leverage Preference Optimization (PO) (Maini et al.

Zhang., Z.; Bao, C.; Pan., X.; Chang, C.; Igarashi, T.; and Zhang, G. p. 0

Where the paper says it

However, training data collected from such sources may contain personally identifiable and/or sensitive information, raising significant privacy concerns, especially with the added complexity of visual signals. For instance, processing visual inputs in VLMs can unintentionally disclose information such as location cues, identity of individuals in the background, etc (Zhang et al.

2025. Through the Lens of Privacy: Exploring Privacy Protection in Vision-Language Model Interactions on Smart Glasses. p. 0
Where the paper says it

Our focus here is on VLM unlearning under practical privacy considerations. In VLMs, sensitive/private information may be present in the vision component, language component, or both.

--- Page 1 ---

Cross-Modal Attention Guided Unlearning in Vision-Language Models is a lightweight and efficient VLM unlearning framework that utilizes crossmodal attention to guide the encoding of unlearning information in visual tokens of low importance for relevant queries. It is practical and effective because it utilizes external modules to encode unlearning information in visual tokens of low importance for relevant queries. p. 0

Where the paper says it

We explore the role of visual tokens for output generation in VLMs using crossmodal attention and utilize it to formulate Cross-Modal Attention Guided Unlearning (CAGUL), a lightweight and efficient VLM unlearning framework. In contrast to computationally expensive model finetuning methods, CAGUL utilizes external modules to encode unlearning information in visual tokens of low importance for relevant queries.

Pruning redundant reasoning tokens

The method leverages the relationship between visual and text tokens represented as cross-modal attention scores computed between the two modes of input. Kaduri, Bagon, and Dekel (2024) empirically demonstrated that VLMs extract fine-grained details and attributes from visual tokens in a spatially localized manner and compress visual information into a small subset of highly attended tokens. p. 0

Where the paper says it

Our method is motivated by the relationship between visual and text tokens represented as cross-modal attention scores computed between the two modes of input. Kaduri, Bagon, and Dekel (2024) empirically demonstrated that VLMs extract fine-grained details and attributes from visual tokens in a spatially localized manner and compress visual information into a small subset of highly attended tokens.

Based on this observation, we conjecture that encoding information relevant to private or non-private queries in the least attended visual tokens can be useful for effective unlearning. p. 1
Where the paper says it

that of prompting with all image tokens. Based on this observation, we conjecture that encoding information relevant to private or non-private queries in the least attended visual tokens can be useful for effective unlearning.

How the training signal is built

The CAGUL framework trains two main components: a discriminator Cϕ and a visual token encoder Fψ. The discriminator is trained with the classification objective Lbce = EZv log pCϕ(lZv). p. 3

Where the paper says it

Our CAGUL framework has two trainable modules, the discriminator Cϕ and the visual token encoder Fψ. The discriminator is trained with the following classification objective Lbce = EZv log pCϕ (l|Zv) (10) The encoder is trained with specific unlearning objectives defined on the forget and retain sets separately.

The encoder is trained with specific unlearning objectives defined on the forget and retain sets separately. Generally, the unlearning objective for retain samples is to maintain the base model’s predictive utility formulated as Lr = E(V,X,Y)∼Dr EY t log pθ(Y tV, X, Y <t). p. 3
Where the paper says it

Generally, the unlearning objective for retain samples is to maintain the base model’s predictive utility. In CAGUL, we define this objective as the standard causal language modeling loss for for all samples in Dr, including the non-private samples for images appearing in the forget set formulated as Lr = E(V,X,Y )∼Dr EY t log pθ(Y t |V, X, Y <t) (11) In the VQA task, the forget objective can be interpreted as obtaining low utility for samples in the forget set Df .

In the VQA task, the forget objective is realized by obtaining low utility for samples in the forget set Df using Preference Optimization (PO) to align outputs with refusal responses. p. 3
Where the paper says it

In CAGUL, we define this objective as the standard causal language modeling loss for for all samples in Dr, including the non-private samples for images appearing in the forget set formulated as Lr = E(V,X,Y )∼Dr EY t log pθ(Y t |V, X, Y <t) (11) In the VQA task, the forget objective can be interpreted as obtaining low utility for samples in the forget set Df . We leverage Preference Optimization (PO) (Maini et al.

Where the gains come from

The primary gains are observed in forgetting quality, where CAGUL Rouge and EM scores on Df demonstrate effective forgetting compared to Finetune. Additionally, CAGUL successfully preserves the retain utility of the finetuned base model. p. 6

Where the paper says it

We report retain metrics on the full retain set Dr and the non-private subset Dnp corresponding to the individuals in Df . We observe that CAGUL successfully preserves the retain utility of the finetuned base model.

The method does not require finetuning any component of the VLM, which ensures that the pre-trained knowledge for general tasks is not manipulated. p. 6
Where the paper says it

Our results show that CAGUL retains the performance of the finetuned base model on both MME and POPE datasets, whereas finetuning baselines significantly degrade in performance. We emphasize that our method does not require finetuning any component of the VLM, which ensures that the pretrained knowledge for general tasks is not manipulated, provided that the classifier can accurately distinguish images in the forget set from retain or general evaluation sets.

CAGUL achieves desirable trade-offs between forget quality and utility comparable to baseline finetuning methods while saving on computational efficiency. p. 6
Where the paper says it

5-VL-7B-Instruct. Overall, our empirical results show that CAGUL achieves desirable trade-offs between forget quality and utility comparable to baseline finetuning methods while saving on computational efficiency.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Cross-Modal Attention Guided Unlearning (CAGUL) allows for unlearning in vision-language models by utilizing a discriminator to identify images in the forget set and a visual token encoder to embed unlearning information in visual tokens having low correlation with the query prompt. This enables removal of sensitive information without altering pre-trained model parameters or incurring retraining costs.

  2. The system can perform subset forgetting, where a user may want to forget a subset of their attributes available to the model rather than forgetting all training data, as formulated in the problem setup for VLM unlearning.

  3. The framework can leverage cross-modal attention scores between visual and text modalities to select k least attended visual tokens with respect to the input text, guiding where unlearning objectives are encoded. This suggests that encoding information relevant to private or non-private queries in the least attended visual tokens can be useful for effective unlearning.

  4. The AI system can distinguish between forget set and retain set images using a discriminator module Cphi, which determines whether a given visual input is paired with a query about private data, formalized as l = Cphi(Zv).

  5. The system can generate refusal responses for forget queries by leveraging Preference Optimization (PO) to align the language model’s outputs with refusal responses instead of the ground truth when training on the forget set Df.

  6. The unlearning objective is formulated as a joint loss L = Lbce + Lf + Lr, balancing binary classification loss for the discriminator, preference optimization loss for forget queries, and standard causal language modeling loss for retain samples.

Abstract

Vision-Language Models (VLMs) have demonstrated remarkable capabilities in multi-modal understanding for tasks such as Visual Question Answering (VQA), which requires models to infer outputs based on visual and textual context simultaneously. However, models may memorize some private and/or sensitive information during training and regurgitate it in inference. Recently, machine unlearning has been leveraged to address such leakage of private data in LLMs. VLMs add a layer of complexity to this process, as the visual context in the query may additionally contain sensitive information. To address this issue, we explore unlearning for VLMs, specifically for the VQA task. We explore the role of visual tokens for output generation in VLMs using cross-modal attention and utilize it to formulate Cross-Modal Attention Guided Unlearning (CAGUL), a lightweight and efficient VLM unlearning framework. In contrast to computationally expensive finetuning methods, CAGUL utilizes external modules to encode unlearning information in visual tokens of low importance for relevant queries. Empirical results show that the transformed visual tokens prevent leakage and retain reference model behavior to achieve desirable trade-offs in forget quality and model utility without altering pretrained model parameters.

Sources

Related papers