Visual pattern-based watermarking for large language model generated text

Shariq Bashir

doi:10.1016/j.neucom.2026.133429

Return

Visual pattern-based watermarking for large language model generated text

2026-03-23

PRE

Shariq Bashir *

DOI：10.1016/j.neucom.2026.133429

Original

Original request for help

Save

Abstract

En 中文

LLM watermarking embeds hidden signals into generated text to enable reliable identification of LLM-generated outputs. Traditional probabilistic watermarking introduces a token-level statistical bias and detects it using threshold-based hypothesis tests. While effective and interpretable in automated settings, such methods provide limited explanatory evidence to a human user about the degree to which the watermark is present in the text and which parts of the text were produced by the LLM. Such visual and interpretable evidence is especially important in forensic and legal cases, where a simple yes/no answer (watermark present or not present) is often insufficient. This paper proposes a probabilistic pattern-based watermarking approach that embeds a structured binary image pattern into LLM-generated text via key-dependent token sampling bias. During detection, the watermark can be recovered as a two-dimensional visual pattern. This enables visual inspection of watermark presence in addition to standard threshold-based decision criteria, without compromising automated detection. Experiments on OPT-1.3B, OPT-2.7B, and LLaMA-7B demonstrate that the proposed approach achieves detection performance comparable to baseline probabilistic watermarking under both clean and adversarial scenarios, while preserving text fluency and semantic fidelity. Human evaluation study further shows that the recovered patterns remain visually recognizable under different adversarial scenarios, illustrating the potential for complementary human-verifiable evidence.

Keywords:

watermarking

large language models

visual pattern

text generation

forensic analysis