Protection methods, terminal devices, and storage media for model weight parameters

By dividing and encrypting the model weight parameters in layers and dynamically selecting necessary parts for local decryption, the problem of low security in traditional model weight protection methods is solved, and high security of model weights is achieved throughout the entire process of storage and inference.

CN121212349BActive Publication Date: 2026-05-26BEIJING YINTUO ZHIAN TECHNOLOGY CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
BEIJING YINTUO ZHIAN TECHNOLOGY CO LTD
Filing Date
2025-09-26
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

Traditional model weight protection methods, while ensuring the normal operation of inference tasks, struggle to achieve full-process security protection for model weights. Static encryption schemes make weights vulnerable to unauthorized acquisition, resulting in low security.

Method used

By dividing the model weight parameters into layers and implementing layered encryption, the necessary parts are dynamically selected for partial decryption to meet the requirements of the inference task, and the system is restored to the encrypted state after the task is completed, thus avoiding the occurrence of the model weights in plaintext.

Benefits of technology

While ensuring the normal execution of inference tasks, it effectively prevents model weight leakage or illegal restoration, thus improving the security of model weights throughout the entire storage and inference process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121212349B_ABST
    Figure CN121212349B_ABST
Patent Text Reader

Abstract

This invention relates to the field of model parameter encryption and discloses a method, terminal device, and storage medium for protecting model weight parameters. The method for protecting model weight parameters includes: dividing the model weight parameters to be protected into multiple weight levels; performing layered encryption on the multiple weight levels to obtain layered encrypted model weights; if an inference task requirement is detected, dynamically selecting necessary parts of the model weights for partial decryption to obtain plaintext local model weight parameters, and responding to the inference task requirement based on the plaintext local model weight parameters; if the inference task requirement is met, restoring the plaintext local model weight parameters to their encrypted state. This invention improves the security of model weights throughout the entire storage and inference process.
Need to check novelty before this filing date? Find Prior Art