EPO Patent Application: Method for Encrypting Neural Network Model Operators and Data

ChangeBridge: EPO Bulletin - AI & Computing (G06N)

Published March 18th, 2026

Detected March 23rd, 2026

Summary

The European Patent Office has published a patent application detailing a method for encrypting neural network model operators and data. The application, filed by XG TECH PTE. LTD., aims to improve the security of neural networks while reducing decryption time.

View original document View source feed page

What changed

This document is a patent application published by the European Patent Office (EPO) concerning a method for encrypting neural network model operators and data. The invention, assigned to XG TECH PTE. LTD., focuses on determining which operators and data volumes within a neural network require encryption, and then encrypting specific subsequences of weights based on encryption modes and positional information. The stated goal is to enhance the safety of neural network models while decreasing the duration of decryption.

As a patent application, this document does not impose direct compliance obligations on regulated entities. However, it represents a development in AI and cybersecurity technology. Companies involved in developing or deploying neural networks, particularly those handling sensitive data or intellectual property, may find the described encryption techniques relevant for their internal R&D or for understanding potential future industry standards or competitive landscapes.

Source document (simplified)

← EPO Patent Bulletin

METHOD AND APPARATUS FOR ENCRYPTING MODEL, ELECTRONIC DEVICE, AND COMPUTER-READABLE STORAGE MEDIUM

Search Report EP4685681A3 Kind: A3 Mar 18, 2026

Applicants

XG TECH PTE. LTD.

Inventors

CAO, Feixiang, MAO, Yunjie, LU, Yawei, LIAO, Ruochen, LI, Mengchen

Abstract

Disclosed are a method and apparatus for encrypting a model, an electronic device, and a computer-readable storage medium. The method includes: determining at least one to-be-encrypted operator in a neural network model; determining a to-be-encrypted data volume corresponding to a respective to-be-encrypted operator of the at least one to-be-encrypted operator based on a data volume adjustment coefficient and a total data volume corresponding to a type of the respective to-be-encrypted operator; determining information on a position of a to-be-encrypted target subsequence of weights in a sequence of weights corresponding to the respective to-be-encrypted operator based on the to-be-encrypted data volume corresponding to the respective to-be-encrypted operator and a number of subsequences of weights corresponding to the respective to-be-encrypted operator; and encrypting the to-be-encrypted target subsequence of weights based on a mode of encryption of the respective to-be-encrypted operator and the information on the position of the to-be-encrypted target subsequence of weights. Using this disclosure, it is enabled to lower the duration of decryption of the neural network model while still guaranteeing safety of the neural network model.

IPC Classifications

G06F 21/60 20130101AFI20260206BHEP G06N 3/063 20230101ALI20260206BHEP G06N 3/08 20230101ALI20260206BHEP G06N 3/04 20230101ALI20260206BHEP G06N 3/045 20230101ALI20260206BHEP G06N 3/082 20230101ALI20260206BHEP G06N 3/084 20230101ALI20260206BHEP

Designated States

AL, AT, BE, BG, CH, CY, CZ, DE, DK, EE, ES, FI, FR, GB, GR, HR, HU, IE, IS, IT, LI, LT, LU, LV, MC, ME, MK, MT, NL, NO, PL, PT, RO, RS, SE, SI, SK, SM, TR

View original document →