Privacy Leakage Across Vision and Language Modalities

31 concepts20 papers1 proposal

Traditional privacy attacks on machine learning models operate within a single modality. In a standard image classification model, a Membership Inference Attack (MIA) evaluates whether a specific sample was in the training set by looking at the model's confidence scores or loss values. If the model outputs an exceptionally high confidence score for a specific image, it likely memorized it during training.

However, the rise of multimodal architectures—such as Vision-Language Models (VLMs) and Multimodal Retrieval-Augmented Generation (MRAG)—fundamentally shifts this dynamic. These models map different data types into a shared semantic space. This cross-modal alignment introduces a unique vulnerability: cross-modal privacy leakage. An attacker no longer needs direct access to a model's internal image representations or classification scores to compromise visual privacy. Instead, they can exploit the text generated by the model to infer sensitive details about the visual inputs, or vice versa.

The Mechanics of Cross-Modal Membership Inference

Why does cross-modal membership inference work? The core intuition lies in how generative models bridge modalities. In an image captioning or visual question-answering task, the model does not output a simple classification probability. Instead, it generates a sequence of text tokens based on a visual input.

To bypass the lack of traditional confidence scores in generative sequence tasks, Pingyi Hu et al. developed the M$^4$I framework. The key insight is that a VLM will exhibit a tighter alignment between its input modality and its generated output modality if the input was seen during training.

This alignment can be quantified through a "re-feeding" technique formalized by Elias Abad Rocamora et al. in their VL-MIA benchmark. Consider an attacker trying to determine if a private image was used to train a VLM:

  1. The attacker feeds the target image into the VLM and allows the model to generate a textual description.
  2. The attacker then re-feeds both the original image and the generated text description back into the model.
  3. By analyzing the resulting logit distributions and token probabilities, the attacker extracts a membership signal.

If the model displays unusually high confidence or low entropy when predicting the text tokens given that specific image, the image is highly likely to be a training member.

Furthermore, Yuke Hu et al. demonstrated in Membership Inference Attacks Against Vision-Language Models that this vulnerability can be exploited even in strict black-box settings by shifting the focus to set-level analysis and leveraging the model's temperature sensitivity. By perturbing the generation temperature, attackers can expose the underlying stability of the model's outputs, revealing membership signals embedded deep within the text generation process without needing direct access to the latent space.

Why Contrastive Alignment Breeds Vulnerability

To understand why these cross-modal signals are so pronounced, we must look at how multimodal encoders are trained. Models like CLIP rely on contrastive learning to pull embeddings of matching image-text pairs closer together while pushing mismatching pairs apart.

In the single-modality space, Hongbin Liu et al. showed with EncoderMI that overfitted contrastive encoders produce highly similar feature vectors for different augmentations of the same training image. When extended to cross-modal settings, this means that an overfitted encoder creates an artificially tight geometric bond between an image and its corresponding text representation. The model essentially carves out a highly specific region in the latent space for that training pair.

When the model encounters a training image during inference, its latent representation is pulled strongly toward the exact text tokens it was paired with during training. This creates the very logit discrepancies and alignment anomalies that cross-modal MIAs exploit.

Indirect Data Leakage and Visual Memorization

Privacy leakage in multimodal models is not limited to binary membership checks; it can result in the complete extraction of sensitive training data. Because these models are trained to translate visual information into text, their language generation components can act as a backdoor, leaking the underlying visual training data that the vision encoder memorized.

Francesco Pinto et al. exposed this phenomenon in Extracting Training Data From Document-Based VQA Models. They discovered that models like Donut and Pix2Struct can memorize sensitive text embedded within training images. Remarkably, during inference, if a user asks about a specific piece of information that has been redacted or visually obscured in the input image, the model can still retrieve the exact text from its weights and output it.

A parallel vulnerability exists in Multimodal Retrieval-Augmented Generation (MRAG) systems. While RAG is often deployed to keep data private by storing it in an external database rather than within model weights, Jiankun Zhang et al. demonstrated in Beyond Text: Unveiling Privacy Vulnerabilities in Multi-modal Retrieval-Augmented Generation that MRAG systems are highly susceptible to indirect data leakage.

Using compositional structured prompt attacks, an attacker can manipulate a Large Multimodal Model to query an external database and leak sensitive non-textual assets like medical CT scans or private signatures. This leakage manifests indirectly: even if the system is restricted from outputting raw media files, the model will generate high-fidelity, detailed textual descriptions of the retrieved image or audio, effectively exfiltrating the private data across the modality boundary.

The Multimodal Defense Gap

The cross-modal nature of these vulnerabilities explains why traditional, single-modality privacy defenses fall short. For instance, differential privacy (DP) frameworks have proven highly effective for text alignment, as noted by Da Yu et al. in Privacy-Preserving Instructions for Aligning Large Language Models. However, adapting these guarantees to the complex, high-dimensional distributions of visual data remains a significant challenge, as highlighted by Haoxiang Wang et al. in Synthesize Privacy-Preserving High-Resolution Images via Private Textual Intermediaries.

Similarly, client-side embedding obfuscation techniques like OSNIP (Zeyu Ma et al.) leverage geometric null spaces to hide text inputs from cloud servers, but it remains an open question whether the complex architectures of vision-language models possess the same geometric properties to allow visual obfuscation without destroying utility. Because a vulnerability in one modality can be exploited through another, securing multimodal systems requires holistic defenses that explicitly account for the semantic bridges connecting vision, audio, and language.

Key Foundational Work

M$^4$I: Multi-modal Models Membership Inference (2022) — contributed 2 concept elements to this direction.
EncoderMI: Membership Inference against Pre-trained Encoders in Contrastive Learning (2021) — contributed 1 concept elements to this direction.
Quantifying and Mitigating Privacy Risks of Contrastive Learning (2021) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cross_modal_membership_inference_9bd6b762,
  title = {cross-modal membership inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-modal-membership-inference/rdr_9bd6b762ecc44bfcbc95a77778907b46},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_cross_modal_membership_inference_9bd6b762_v1,
  title = {cross-modal membership inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-modal-membership-inference/rdr_9bd6b762ecc44bfcbc95a77778907b46/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake