Federated AI Model Updates With Sensitive Parameter Separation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In federated learning scenarios, data distribution heterogeneity and privacy concerns hinder the effective aggregation of models across clients, leading to suboptimal performance and increased communication costs, especially in industrial applications like visual quality control.

Innovation Solution

A method that identifies and separates sensitive model parameters from non-sensitive ones, allowing clients to update their models selectively with a global model while retaining privacy, by checking gradients and accuracy thresholds, and only sharing non-sensitive parameters.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If federated learning aggregates models from multiple clients to improve overall model performance, then model accuracy is improved, but data privacy is compromised and communication costs increase

Engineering Contradiction:
Improvemodel accuracyVSAvoiddata privacy
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent segments model parameters into sensitive and non-sensitive categories based on their contribution to model accuracy. Only non-sensitive parameters are shared with the server for aggregation, while sensitive parameters remain local. This segmentation allows model improvement through federated learning without compromising data privacy, as the server never receives or processes sensitive client data.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extracts and identifies sensitive model parameters that should not be shared, separating them from non-sensitive parameters. By taking out the sensitive parameters from the sharing process, the system maintains privacy while still enabling beneficial knowledge transfer through aggregation of non-sensitive parameters across clients.

Inventive Principle:
Principle #2Taking out (Extraction)

2Measurement precision

If all model parameters are shared in federated learning to improve model performance, then model accuracy is improved, but communication costs increase

Engineering Contradiction:
Improvemodel accuracyVSAvoidcommunication costs
Core Design Contradiction:
Measurement precisionVSLoss of energy

Solution Approach 1:

Instead of sharing all model parameters (excessive action), the patent applies partial action by sharing only the non-sensitive parameters that are necessary for model improvement. This selective sharing reduces communication bandwidth and energy consumption while still achieving the goal of improving model accuracy through federated learning.

Inventive Principle:
Principle #16Partial or excessive action

3Stability of the object's composition

If a global model is trained for all clients to standardize performance, then model consistency is improved, but individual client performance deteriorates due to non-IID data distribution

Engineering Contradiction:
Improvemodel consistencyVSAvoidindividual client performance
Core Design Contradiction:
Stability of the object's compositionVSMeasurement precision

Solution Approach 1:

The patent applies local quality by allowing each client to maintain and optimize their own model parameters locally while still benefiting from the global model. Clients can adapt the global model to their specific non-IID data distributions by keeping sensitive parameters local, thus achieving both model consistency through global aggregation and individual performance optimization through local adaptation.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4455948B1Computer-implemented method and system for operating a technical device
Publication Date: 2026.01.14 SIEMENS AG
  • EP4455948B1 patent drawingFigure 1~2
  • EP4455948B1 patent drawingFigure 3
  • EP4455948B1 patent drawing

AI summary

A computer-implemented method for operating a technical device (TD) with an artificial intelligence-based model by a client (C1-C3) of a client-server system, wherein each client (C1-C3) stores a client model (CM1-CM3) for operating a connected technical device (TD1), and the following steps are performed: a) providing the global model (GM) based on federated learning to a client (C1), b) the client (C1-C3) checking whether a sensitive model parameter is stored in memory (MEM) that is not included in the global model (GM), and if so, aggregating the provided global model (GM) with the sensitive model parameter, and updating the client model (CM1) with the provided global model (GM), c) providing a reference dataset, d) calculating an initial accuracy of the client model (CM1) using the Reference data set,e) Determine the gradients of the model parameters of the client model (CM1) and determine a selected gradient of the model parameters of the client model (CM1), f) Remove the model parameter corresponding to the selected gradient from the client model (CM1), g) Calculate a second accuracy of the client model (CM1) using the reference dataset, h) Check if the second accuracy is lower than the first accuracy; if so, designate the at least one model parameter corresponding to the at least one selected gradient as the sensitive parameter and provide the client model (CM1) to the server (S), i) Update the global model, j) Operate the technical device (TD1) with the client model (CM1) and the sensitive parameter.