Beyond Text: Unveiling Privacy Vulnerabilities in Multi-modal Retrieval-Augmented Generation

2025/05/20 by Jiankun Zhang, Zhang, Jiankun, Shenglai Zeng +11 · 1 citation
Computer Science · Social Sciences · #Access Control and Trust #Computation and Language (cs.CL) #Cryptography and Security (cs.CR) #FOS: Computer and information sciences #Privacy-Preserving Technologies in Data #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2505.13957

openalex publication_date 2025/05/20 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Multimodal Retrieval-Augmented Generation (MRAG) systems enhance LMMs by integrating external multimodal databases, but introduce unexplored privacy vulnerabilities. While text-based RAG privacy risks have been studied, multimodal data presents unique challenges. We provide the first systematic analysis of MRAG privacy vulnerabilities across vision-language and speech-language modalities. Using a novel compositional structured prompt attack in a black-box setting, we demonstrate how attackers can extract private information by manipulating queries. Our experiments reveal that LMMs can both directly generate outputs resembling retrieved content and produce descriptions that indirectly expose sensitive information, highlighting the urgent need for robust privacy-preserving MRAG techniques.

Citations

• Empowering GraphRAG with Knowledge Filtering and Integration
• Towards Context-Robust LLMs: A Gated Representation Fine-tuning Approach
• Unveiling Privacy Risks in LLM Agent Memory
• Retrieval-Augmented Generation with Graphs (GraphRAG)
• Towards Knowledge Checking in Retrieval-augmented Generation: A Representation Perspective
• Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench
• MMed-RAG: Versatile Multimodal RAG System for Medical Vision Language Models
• MiniCPM-V: A GPT-4V Level MLLM on Your Phone
• Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech Generation
• MMedAgent: Learning to Use Medical Tools with Multi-modal Agent
• Let's Go Real Talk: Spoken Dialogue Model for Face-to-Face Conversation
• Noise Masking Attacks and Defenses for Pretrained Speech Models
• The Good and The Bad: Exploring Privacy Issues in Retrieval-Augmented Generation (RAG)
• Copyright Protection in Generative AI: A Technical Perspective
• Can Language Models be Instructed to Protect Personal Information?
• Large Language Models
• REVEAL: Retrieval-Augmented Visual-Language Pre-Training with Multi-Source Multimodal Knowledge Memory
• Large-scale Contrastive Language-Audio Pretraining with Feature Fusion and Keyword-to-Caption Augmentation
• Retrieval Augmented Visual Question Answering with Outside Knowledge
• MuRAG: Multimodal Retrieval-Augmented Generator for Open Question Answering over Images and Text
• Re-Imagen: Retrieval-Augmented Text-to-Image Generator
• Extracting Targeted Training Data from ASR Models, and How to Mitigate It
• Align before Fuse: Vision and Language Representation Learning with Momentum Distillation
• Retrieval Augmentation Reduces Hallucination in Conversation
• Learning Transferable Visual Models From Natural Language Supervision
• Extracting Training Data from Large Language Models
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• Generalization through Memorization: Nearest Neighbor Language Models
• VQA: Visual Question Answering
• Distinctive Image Features from Scale-Invariant Keypoints

Cited by

• Safe-LLaVA: A Privacy-Preserving Vision-Language Dataset and Benchmark for Biometric Safety

Related

• MrM: Black-Box Membership Inference Attacks against Multimodal RAG Systems
• The Good and The Bad: Exploring Privacy Issues in Retrieval-Augmented Generation (RAG)
• RAG Security and Privacy: Formalizing the Threat Model and Attack Surface
• Fine-Grained Privacy Extraction from Retrieval-Augmented Generation Systems via Knowledge Asymmetry Exploitation
• Safeguarding Privacy of Retrieval Data against Membership Inference Attacks: Is This Query Too Close to Home?