Graph Neural Network Training via Federated Multi-Party Computation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing graph neuronal networks face challenges in processing large-scale, complex, and dynamic data from industrial systems with multiple premises, each operated by multiple parties with limited trust in each other. These systems struggle with heterogeneous graphs, intricate relationships, and scalability, while also requiring decentralized data processing to maintain privacy.

Innovation Solution

A computer-implemented method is proposed that trains a graph neuronal network using privacy-preserving multi-party computation and decentralized federated learning. This method preprocesses operational graph data using privacy-preserving techniques and applies end-to-end privacy preservation throughout the training process, ensuring that no party can make inferences about other parties' data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If data centralization is used to train graph neuronal networks, then model training efficiency is improved, but data privacy and security are compromised

Engineering Contradiction:
Improvemodel training efficiencyVSAvoiddata privacy security
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent segments the centralized training process into distributed federated learning across multiple premises. Each premise trains local graph neuronal networks on its own data, then shares only model updates (not raw data) with a central coordinator. This segmentation enables both distributed privacy preservation and coordinated model training.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces cryptographic intermediaries (secure multi-party computation protocols, homomorphic encryption, or trusted execution environments) that enable collaborative model training without direct data sharing. These intermediaries act as mediators that allow computations on encrypted data, preserving privacy while enabling productivity.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If decentralized federated learning is used to preserve privacy, then data privacy is improved, but model training complexity increases

Engineering Contradiction:
Improvedata privacyVSAvoidmodel training complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

Each premise performs self-service by training its own local graph neuronal network on its operational data without requiring centralized data collection. The local training process is autonomous, and only model parameter updates are shared, reducing the complexity burden on any single entity while maintaining privacy.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent employs universal cryptographic primitives (such as homomorphic encryption or secure multi-party computation protocols) that can be applied across different premises and data types. These multi-functional cryptographic tools handle various privacy-preserving operations uniformly, reducing overall system complexity despite the decentralized architecture.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If multiple parties process operational graph data, then collaborative inference accuracy is improved, but privacy leakage risk increases

Engineering Contradiction:
Improveinference accuracyVSAvoidprivacy leakage risk
Core Design Contradiction:
Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent applies local quality by allowing each party to process and contribute only its local operational graph data to the collaborative model training. Each premise's data characteristics and quality are preserved locally, and the aggregated model benefits from diverse local contributions without any single party's data being exposed to others.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent transforms raw operational data into model parameter updates through the training process. Instead of sharing sensitive raw data, parties exchange transformed parameter representations that capture learning insights without revealing underlying operational details. This parameter transformation maintains inference accuracy while reducing privacy leakage risk.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4542450A1Method of training a graph neuronal network to derive an inference about operations in an industrial system com-prising a plurality of premises
Publication Date: 2025.04.23 SIEMENS AG
  • EP4542450A1 patent drawingFigure 1
  • EP4542450A1 patent drawingFigure 2~4
  • EP4542450A1 patent drawingFigure 5~7

AI summary

The proposed method comprises: preprocessing respective operational graph data (111-114) related to the technical operations and collected in each of the premises (100, 200) by each of a plurality of parties operating the respective premise (100, 200) using a privacy-preserving multi-party computation technique to obtain preprocessed operational graph data (211-214); and using decentralized federated learning with a privacy-preserving model update method to train the graph neuronal network (3) to provide, in response to the preprocessed operational graph data (211-214) of each of the premises (100, 200) being input, the technical inference (4) as an output. An additional end-to-end privacy preservation technique is applied throughout. The method ensures privacy of operations of each of the parties to a level sufficient for industrial use.