Reproduce “VLA-Mark: A cross modal watermark for large vision-language alignment model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shuliang Liu, Qi Zheng, Jesse Jiaxi Xu, Yibo Yan, Junyan Zhang, He Geng, Aiwei Liu, Peijie Jiang and 3 more
VLA-Mark introduces a cross-modal watermarking framework for Vision-Language Alignment Models (VLAMMs), embedding detectable watermarks into generated text while preserving multimodal coherence and semantic fidelity. The approach achieves over 99.8% AUROC detection and lower perplexity compared to baselines across multiple VLAMMs, along with improved robustness against adversarial attacks.
Key insights
- Leveraging a VLA model's inherent cross-modal alignment allows for dynamic, content-aware watermarking that directly maintains multimodal coherence and semantic fidelity.
- Fusing multiscale semantic saliency metrics provides a robust mechanism to identify and protect vision-critical tokens, adapting watermark strength based on their visual importance.
- Dynamically adjusting watermark intensity based on generation entropy enables a flexible trade-off between detectability and output quality, adapting injection to the model's confidence level.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.