Scholarlink

A novel zero-knowledge proof (ZKP) paradigm enables verifiable language model adaptation without exposing proprietary parameters or datasets. Traditional verification requires full disclosure, creating privacy risks in sensitive domains like healthcare and finance. Our approach integrates polynomial commitments with an optimized proof scheme using Kate-Zaverucha-Goldberg (KZG) commitments to ensure adaptation integrity while preserving confidentiality. Implementation with a GPT-2 model fine-tuned on medical data achieved 46.2% loss reduction while generating cryptographic proofs in 84.44 s, with 36 MB proof size and 0.303 compression ratio. The system produced 124,503 commitments with average polynomial degree of 10.0. Total runtime was 142.57 s (adaptation: 58.13 s; proof generation: 84.44 s), with 3.36 GB memory usage. The workflow includes capability evaluation, fine-tuning, re-evaluation, polynomial conversion, commitment generation, and proof storage. While challenges remain in proof generation costs and scalability, this framework advances privacy-preserving AI verification for sensitive applications.

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

Zero-Knowledge Enhanced Adaptation Learning (ZEAL): A Framework for Verifiable Model Adaptation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dhrumil Panchal,
  • Hrushik Mehta,
  • Pragati Shetty,
  • Pratik Kanani,
  • Darshana Sankhe,
  • Rashmi Kumar

摘要

A novel zero-knowledge proof (ZKP) paradigm enables verifiable language model adaptation without exposing proprietary parameters or datasets. Traditional verification requires full disclosure, creating privacy risks in sensitive domains like healthcare and finance. Our approach integrates polynomial commitments with an optimized proof scheme using Kate-Zaverucha-Goldberg (KZG) commitments to ensure adaptation integrity while preserving confidentiality. Implementation with a GPT-2 model fine-tuned on medical data achieved 46.2% loss reduction while generating cryptographic proofs in 84.44 s, with 36 MB proof size and 0.303 compression ratio. The system produced 124,503 commitments with average polynomial degree of 10.0. Total runtime was 142.57 s (adaptation: 58.13 s; proof generation: 84.44 s), with 3.36 GB memory usage. The workflow includes capability evaluation, fine-tuning, re-evaluation, polynomial conversion, commitment generation, and proof storage. While challenges remain in proof generation costs and scalability, this framework advances privacy-preserving AI verification for sensitive applications.