Auditing M-LLMs for Privacy Risks: A Synthetic Benchmark and Evaluation Framework

2025-11-05

AI Summary

Social media multimodal content implicitly contains sensitive personal information, and multimodal large language models (M-LLMs) may inadvertently infer such attributes—posing a novel privacy risk. To address this, we introduce PRISM, the first privacy-risk assessment benchmark for M-LLMs: a fine-grained synthetic dataset covering 12 sensitive attribute categories and simulating user dynamic history, coupled with an automated, multi-agent reasoning–based evaluation framework. Leveraging LLM-agentic workflows to generate data aligned with prior distributions and domain-specific pipelines, we evaluate six state-of-the-art M-LLMs—including Qwen and GPT-4o. Experiments reveal that current M-LLMs significantly outperform humans in inferring sensitive attributes, underscoring the urgency of mitigation strategies. PRISM is publicly released on Hugging Face to support reproducible, scalable privacy evaluation.

Technology Category

• Machine Learning: Privacy
• Natural Language Processing: Safety and Robustness
• Philosophy and Ethics of AI: Privacy & Security

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Security and Privacy: Large-scale security measurements
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Recent advances in multi-modal Large Language Models (M-LLMs) have demonstrated a powerful ability to synthesize implicit information from disparate sources, including images and text. These resourceful data from social media also introduce a significant and underexplored privacy risk: the inference of sensitive personal attributes from seemingly daily media content. However, the lack of benchmarks and comprehensive evaluations of state-of-the-art M-LLM capabilities hinders the research of private attribute profiling on social media. Accordingly, we propose (1) PRISM, the first multi-modal, multi-dimensional and fine-grained synthesized dataset incorporating a comprehensive privacy landscape and dynamic user history; (2) an Efficient evaluation framework that measures the cross-modal privacy inference capabilities of advanced M-LLM. Specifically, PRISM is a large-scale synthetic benchmark designed to evaluate cross-modal privacy risks. Its key feature is 12 sensitive attribute labels across a diverse set of multi-modal profiles, which enables targeted privacy analysis. These profiles are generated via a sophisticated LLM agentic workflow, governed by a prior distribution to ensure they realistically mimic social media users. Additionally, we propose a Multi-Agent Inference Framework that leverages a pipeline of specialized LLMs to enhance evaluation capabilities. We evaluate the inference capabilities of six leading M-LLMs (Qwen, Gemini, GPT-4o, GLM, Doubao, and Grok) on PRISM. The comparison with human performance reveals that these MLLMs significantly outperform in accuracy and efficiency, highlighting the threat of potential privacy risks and the urgent need for robust defenses. Dataset available at https://huggingface.co/datasets/xaddh/multimodal-privacy

Problem

• Evaluating privacy risks in multi-modal LLMs from social media content
• Lacking benchmarks for sensitive attribute inference capabilities assessment
• Measuring cross-modal privacy inference in advanced M-LLM systems

Innovation

• Synthetic multi-modal dataset with privacy labels
• Multi-agent framework for enhanced evaluation
• Benchmarking six M-LLMs against human performance