Visible watermarks are commonly used to indicate document ownership, confidentiality, or status. But according to a new paper on arXiv, they can also be repurposed as carriers of adversarial signals. The authors introduce Auto-Regional Adversarial Watermarking (AR-AWM), a model-aware white-box attack aimed at end-to-end vision-language models (VLMs) used for document understanding.

The attack is white-box, meaning the adversary has full access to the model's architecture and parameters. By modifying watermarks in a region-specific way, AR-AWM can inject misleading information that the VLM picks up during document processing. The paper's abstract does not yet specify the attack's success rates or the types of documents tested, but the core idea is that a seemingly benign visual element can become a vector for manipulation.

This work highlights a potential blind spot in end-to-end VLMs: they may trust visible watermarks as harmless metadata, while an attacker can turn those same marks into a tool for confusion. The authors position AR-AWM as a step toward understanding and defending against such watermark-constrained threats in document AI.