Scholarlink

The rapid proliferation of large language models has raised concerns about unauthorized redistribution, uncredited fine-tuning, and the erosion of intellectual property value. Existing watermarking methods often degrade model quality, have limited payload capacity, or fail under post-processing steps such as quantization and fine-tuning. We introduce TransMark, a watermarking framework that embeds a high-entropy binary payload into a model’s parameters while keeping its functional behavior virtually unchanged. Our method identifies pairs of feed-forward neurons that have nearly identical \(L_2\text {-norm}\) but sufficiently divergent directional vectors and encodes bits by swapping the corresponding rows (and columns) without altering layer-wise activations in expectation. A locality-aware search enforces tight similarity thresholds, enabling dense yet imperceptible bit injection. To ensure decoding reliability, the bitstream is protected through repeat-code voting, and a lightweight meta structure stores only minimal auxiliary data. Experiments show that TransMark preserves perplexity and generation quality, adds negligible computational overhead, and remains robust against integer quantization, additive Gaussian noise, and single-step low-rank adaptation.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust Ownership Verification in Large Language Models via Equivalent Neuron Pair Encoding

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zhuorong Chen,
  • Pei-Gen Ye

摘要

The rapid proliferation of large language models has raised concerns about unauthorized redistribution, uncredited fine-tuning, and the erosion of intellectual property value. Existing watermarking methods often degrade model quality, have limited payload capacity, or fail under post-processing steps such as quantization and fine-tuning. We introduce TransMark, a watermarking framework that embeds a high-entropy binary payload into a model’s parameters while keeping its functional behavior virtually unchanged. Our method identifies pairs of feed-forward neurons that have nearly identical \(L_2\text {-norm}\) but sufficiently divergent directional vectors and encodes bits by swapping the corresponding rows (and columns) without altering layer-wise activations in expectation. A locality-aware search enforces tight similarity thresholds, enabling dense yet imperceptible bit injection. To ensure decoding reliability, the bitstream is protected through repeat-code voting, and a lightweight meta structure stores only minimal auxiliary data. Experiments show that TransMark preserves perplexity and generation quality, adds negligible computational overhead, and remains robust against integer quantization, additive Gaussian noise, and single-step low-rank adaptation.