Unveiling Privacy Risks in Multi-modal Large Language Models: Task-specific Vulnerabilities and Mitigation Challenges
2026-06-08 • Cryptography and Security
Cryptography and SecurityArtificial Intelligence
AI summaryⓘ
The authors studied how Multi-modal Large Language Models (MLLMs), which understand both images and text, can accidentally reveal private information. They created a dataset called MM-Privacy to test how these models leak sensitive data from images and memory. Their tests showed that MLLMs are vulnerable to privacy risks in different tasks and highlighted how inconsistent tasks can increase these risks. The authors stress the need for better protections to keep information safe when using MLLMs.
Multi-modal Large Language ModelsPrivacy RisksData LeakageSensitive InformationImage ProcessingText ProcessingDatasetModel EvaluationTask InconsistencyMitigation Strategies
Authors
Tiejin Chen, Pingzhi Li, Kaixiong Zhou, Tianlong Chen, Hua Wei
Abstract
Privacy risks in text-only Large Language Models (LLMs) are well studied, particularly their tendency to memorize and leak sensitive information. However, Multi-modal Large Language Models (MLLMs), which process both text and images, introduce unique privacy challenges that remain underexplored. Compared to text-only models, MLLMs can extract and expose sensitive information embedded in images, posing new privacy risks. We reveal that some MLLMs are susceptible to privacy breaches, leaking sensitive data embedded in images or stored in memory. Specifically, in this paper, we (1) introduce MM-Privacy, a comprehensive dataset designed to assess privacy risks across various multi-modal tasks and scenarios, where we define Disclosure Risks and Retention Risks. (2) systematically evaluate different MLLMs using MM-Privacy and demonstrate how models leak sensitive data across various tasks, and (3) provide additional insights into the role of task inconsistency in privacy risks, emphasizing the urgent need for mitigation strategies. Our findings highlight privacy concerns in MLLMs, underscoring the necessity of safeguards to prevent data exposure. Our dataset and code can be found here.