Abstract:Large Multimodal Language Models (MLLMs) are emerging as one of the foundational tools in an expanding range of applications. Consequently, understanding training-data leakage in these systems is increasingly critical. Log-probability-based membership inference attacks (MIAs) have become a widely adopted approach for assessing data exposure in large language models (LLMs), yet their effect in MLLMs remains unclear. We present the first comprehensive evaluation of extending these text-based MIA methods to multimodal settings. Our experiments under vision-and-text (V+T) and text-only (T-only) conditions across the DeepSeek-VL and InternVL model families show that in in-distribution settings, logit-based MIAs perform comparably across configurations, with a slight V+T advantage. Conversely, in out-of-distribution settings, visual inputs act as regularizers, effectively masking membership signals.
| Comments: | accepted by ESANN 2026 |
| Subjects: | Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2512.03121 [cs.CR] |
| (or arXiv:2512.03121v2 [cs.CR] for this version) | |
| https://doi.org/10.48550/arXiv.2512.03121 arXiv-issued DOI via DataCite |
From: Ziyi Tong [view email]
[v1]
Tue, 2 Dec 2025 14:11:51 UTC (299 KB)
[v2]
Thu, 21 May 2026 10:20:13 UTC (299 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。