Saved in:
Bibliographic Details
Main Authors: Ju, Tianjie, Hua, Yi, Fei, Hao, Shao, Zhenyu, Zheng, Yubin, Zhao, Haodong, Lee, Mong-Li, Hsu, Wynne, Zhang, Zhuosheng, Liu, Gongshen
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2503.01208
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915342667743232
author Ju, Tianjie
Hua, Yi
Fei, Hao
Shao, Zhenyu
Zheng, Yubin
Zhao, Haodong
Lee, Mong-Li
Hsu, Wynne
Zhang, Zhuosheng
Liu, Gongshen
author_facet Ju, Tianjie
Hua, Yi
Fei, Hao
Shao, Zhenyu
Zheng, Yubin
Zhao, Haodong
Lee, Mong-Li
Hsu, Wynne
Zhang, Zhuosheng
Liu, Gongshen
contents Multi-Modal Large Language Models (MLLMs) have exhibited remarkable performance on various vision-language tasks such as Visual Question Answering (VQA). Despite accumulating evidence of privacy concerns associated with task-relevant content, it remains unclear whether MLLMs inadvertently memorize private content that is entirely irrelevant to the training tasks. In this paper, we investigate how randomly generated task-irrelevant private content can become spuriously correlated with downstream objectives due to partial mini-batch training dynamics, thus causing inadvertent memorization. Concretely, we randomly generate task-irrelevant watermarks into VQA fine-tuning images at varying probabilities and propose a novel probing framework to determine whether MLLMs have inadvertently encoded such content. Our experiments reveal that MLLMs exhibit notably different training behaviors in partial mini-batch settings with task-irrelevant watermarks embedded. Furthermore, through layer-wise probing, we demonstrate that MLLMs trigger distinct representational patterns when encountering previously seen task-irrelevant knowledge, even if this knowledge does not influence their output during prompting. Our code is available at https://github.com/illusionhi/ProbingPrivacy.
format Preprint
id arxiv_https___arxiv_org_abs_2503_01208
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Watch Out Your Album! On the Inadvertent Privacy Memorization in Multi-Modal Large Language Models
Ju, Tianjie
Hua, Yi
Fei, Hao
Shao, Zhenyu
Zheng, Yubin
Zhao, Haodong
Lee, Mong-Li
Hsu, Wynne
Zhang, Zhuosheng
Liu, Gongshen
Computer Vision and Pattern Recognition
Computation and Language
Multi-Modal Large Language Models (MLLMs) have exhibited remarkable performance on various vision-language tasks such as Visual Question Answering (VQA). Despite accumulating evidence of privacy concerns associated with task-relevant content, it remains unclear whether MLLMs inadvertently memorize private content that is entirely irrelevant to the training tasks. In this paper, we investigate how randomly generated task-irrelevant private content can become spuriously correlated with downstream objectives due to partial mini-batch training dynamics, thus causing inadvertent memorization. Concretely, we randomly generate task-irrelevant watermarks into VQA fine-tuning images at varying probabilities and propose a novel probing framework to determine whether MLLMs have inadvertently encoded such content. Our experiments reveal that MLLMs exhibit notably different training behaviors in partial mini-batch settings with task-irrelevant watermarks embedded. Furthermore, through layer-wise probing, we demonstrate that MLLMs trigger distinct representational patterns when encountering previously seen task-irrelevant knowledge, even if this knowledge does not influence their output during prompting. Our code is available at https://github.com/illusionhi/ProbingPrivacy.
title Watch Out Your Album! On the Inadvertent Privacy Memorization in Multi-Modal Large Language Models
topic Computer Vision and Pattern Recognition
Computation and Language
url https://arxiv.org/abs/2503.01208