Secure Multi-Party Neural Training With Hidden Intersection IDs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing secure multi-party computing protocols expose customer IDs during private set intersection, compromising data privacy and security in vertical secure multi-party learning.

Innovation Solution

Implement a method where computing participants use secret sharing and oblivious programmable pseudorandom functions to determine and protect intersection IDs, sharing feature values only for AI model training, and delete non-intersection data to enhance security and efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If private set intersection technology is used to calculate intersection of IDs, then data alignment for vertical secure multi-party learning is achieved, but ID privacy is exposed and security deteriorates

Engineering Contradiction:
Improvedata alignment capabilityVSAvoidID privacy protection
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent segments the ID and feature value into separate processing streams. IDs are processed through PSI to determine intersection without exposing plaintext, while feature values are processed separately through secret sharing. This segmentation allows data alignment functionality while preventing ID privacy exposure.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces ciphertext data and secret sharing as intermediary mechanisms. Instead of directly exposing intersection IDs, the system uses ciphertext indicators and shared secrets as intermediaries to convey intersection information without revealing the actual ID values, thus maintaining privacy while enabling data alignment.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If all computing parties know specific IDs in the intersection, then data alignment is completed, but ID privacy of customers is exposed and security becomes poor

Engineering Contradiction:
Improvedata alignment efficiencyVSAvoidID privacy exposure
Core Design Contradiction:
ProductivityVSObject-affected harmful factors

Solution Approach 1:

The patent creates a copy of the intersection identification process using ciphertext data and secret sharing. Instead of directly revealing intersection IDs, the system generates encrypted indicators and shared secret values that can be used to identify intersection records without exposing the actual ID information, thus maintaining privacy while preserving alignment efficiency.

Inventive Principle:
Principle #26Copying

3Reliability

If feature values are shared using secret sharing mode, then data privacy is protected during transmission, but computational complexity increases

Engineering Contradiction:
Improvedata privacy protectionVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies secret sharing selectively only to feature values that need to be protected, rather than applying it universally to all data processing. The secret sharing is implemented locally at each computing party's end for the specific feature value transmission, providing targeted privacy protection with minimized computational overhead.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4701128A1Data processing method and system, and computing device
Publication Date: 2026.02.25 HUAWEI CLOUD COMPUTING TECHNOLOGIES CO LTD
  • EP4701128A1 patent drawingFigure 1~2
  • EP4701128A1 patent drawingFigure 3~5
  • EP4701128A1 patent drawingFigure 6~8

AI summary

A data processing method and system, and a computing device are provided. The method includes: A first computing participant and a second computing participant separately obtain first ciphertext data. The second computing participant obtains, from the first computing participant, a share of a first feature value corresponding to a first ID. The first computing participant obtains second data, where if the first ID is an intersection ID, the second data is a share of a second feature value that is in the second computing participant and that corresponds to the first ID, or if the first ID is not an intersection ID, the second data is a random number. If the first ciphertext data indicates that the first ID is an intersection ID, the first computing participant uses the second data and a share of the first feature value that is held by the first computing participant as training data of a neural network, and the second computing participant uses a share of the first feature value and a share of the second feature value that are held by the second computing participant as training data of the neural network. In the method, security of multi-party computing can be improved.