Leave My Images Alone: Preventing Multi-Modal Large Language Models from Analyzing Images via Visual Prompt Injection
2026-04-10 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionArtificial IntelligenceCryptography and SecurityMachine Learning
AI summaryⓘ
The authors introduce ImageProtector, a tool that slightly changes images in a way people can't easily see but stops certain computer programs called multi-modal large language models (MLLMs) from identifying sensitive information. When these protected images are analyzed, the MLLMs respond by refusing to provide details. The authors tested ImageProtector on several models and datasets, finding it effective. They also looked at some ways to counter this protection but found those methods either reduce accuracy or slow down the models. This work explores how small changes to images can help protect privacy against automated large-scale image analysis.
Multi-modal large language modelsMLLMsImage perturbationPrivacy protectionVisual prompt injectionAdversarial examplesGaussian noiseDiffPureAdversarial trainingOpen-weight models
Authors
Zedian Shao, Hongbin Liu, Yuepeng Hu, Neil Zhenqiang Gong
Abstract
Multi-modal large language models (MLLMs) have emerged as powerful tools for analyzing Internet-scale image data, offering significant benefits but also raising critical safety and societal concerns. In particular, open-weight MLLMs may be misused to extract sensitive information from personal images at scale, such as identities, locations, or other private details. In this work, we propose ImageProtector, a user-side method that proactively protects images before sharing by embedding a carefully crafted, nearly imperceptible perturbation that acts as a visual prompt injection attack on MLLMs. As a result, when an adversary analyzes a protected image with an MLLM, the MLLM is consistently induced to generate a refusal response such as "I'm sorry, I can't help with that request." We empirically demonstrate the effectiveness of ImageProtector across six MLLMs and four datasets. Additionally, we evaluate three potential countermeasures, Gaussian noise, DiffPure, and adversarial training, and show that while they partially mitigate the impact of ImageProtector, they simultaneously degrade model accuracy and/or efficiency. Our study focuses on the practically important setting of open-weight MLLMs and large-scale automated image analysis, and highlights both the promise and the limitations of perturbation-based privacy protection.