Scholarlink

As datasets become critical assets in machine learning, traditional copyright mechanisms fail to address replication and unauthorized use in decentralized environments. This survey categorizes protection approaches into three classes: non-intrusive (ownership detection without data modification), minimally-intrusive (lightweight reversible embeddings), and maximally-intrusive (aggressive alterations like reversible adversarial examples). We analyze their effectiveness, highlight underexplored gaps in text/audio datasets and scalability challenges, and propose unified, ethical solutions for dataset protection in evolving ML ecosystems.

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

Dataset Ownership in the Era of Large Language Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kun Li,
  • Cheng Wang,
  • Minghui Xu,
  • Yue Zhang,
  • Xiuzhen Cheng

摘要

As datasets become critical assets in machine learning, traditional copyright mechanisms fail to address replication and unauthorized use in decentralized environments. This survey categorizes protection approaches into three classes: non-intrusive (ownership detection without data modification), minimally-intrusive (lightweight reversible embeddings), and maximally-intrusive (aggressive alterations like reversible adversarial examples). We analyze their effectiveness, highlight underexplored gaps in text/audio datasets and scalability challenges, and propose unified, ethical solutions for dataset protection in evolving ML ecosystems.