Obfuscation Network Training for Privacy-Preserving ML Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing data anonymization techniques distort original data, making it unusable for machine learning algorithms and fail to conceal unexpected identification information, while personal information protection laws require anonymization without human-recognizable distortion.

Innovation Solution

A method for training an obfuscation network to generate obfuscated data that is different from original data yet recognizable by machine learning models, using a surrogate network trained with obfuscated data to minimize losses and maximize obfuscated data scores, ensuring privacy and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Object-affected harmful factors

If conventional anonymization techniques (blurring, masking) are applied to conceal personal information, then personal information protection is improved, but data utility for machine learning deteriorates due to distortion

Engineering Contradiction:
Improvepersonal information exposureVSAvoiddata utility for machine learning
Core Design Contradiction:
Object-affected harmful factorsVSReliability

Solution Approach 1:

The patent replaces conventional mechanical anonymization methods (blurring, masking) with a neural network-based obfuscation system. The obfuscation network learns to transform data in a way that preserves machine-learning utility while concealing personal information, overcoming the limitations of rule-based anonymization techniques.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the parameters of data transformation from simple geometric operations (blurring radius, masking regions) to learned transformation parameters through neural networks. The obfuscation network adjusts transformation parameters dynamically to maintain both privacy protection and data utility for different data types and machine learning tasks.

Inventive Principle:
Principle #35Parameter changes

2Object-affected harmful factors

If obfuscation network transforms original data to conceal personal information, then privacy protection is improved, but machine learning algorithm performance deteriorates due to data distortion

Engineering Contradiction:
Improvepersonal information exposureVSAvoidmachine learning algorithm accuracy
Core Design Contradiction:
Object-affected harmful factorsVSMeasurement precision

Solution Approach 1:

The patent implements a feedback mechanism where the obfuscation network and surrogate network are trained jointly using loss functions that measure both privacy protection quality and machine learning task performance. The surrogate network provides feedback on how well the obfuscated data maintains utility for the original machine learning task, allowing the obfuscation network to adjust its transformations accordingly.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent performs preliminary training of both the obfuscation network and surrogate network before actual data processing. This preliminary action establishes the transformation relationships and utility preservation characteristics in advance, ensuring that when the system processes real data, the obfuscated output maintains machine learning utility without requiring real-time adjustments that could compromise performance.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If surrogate network is trained with obfuscated data, then data utility is maintained, but training complexity increases due to dual network architecture

Engineering Contradiction:
Improvedata utility for machine learningVSAvoiddual network architecture
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges the training processes of the obfuscation network and surrogate network into a unified joint training framework. By combining the loss functions and training loops, the system optimizes both networks simultaneously, reducing the overall training complexity compared to separate independent training processes while maintaining the benefits of the dual-network architecture.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The surrogate network is designed with multi-functionality, serving both as a utility verification mechanism during obfuscation network training and as a standalone model for processing obfuscated data. This universal design reduces the need for additional separate systems, thereby managing training complexity while maintaining data utility.

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

4Object-affected harmful factors

If iterative training process is used to minimize losses, then obfuscation quality is improved, but training time increases

Engineering Contradiction:
Improvepersonal information exposureVSAvoidtraining time
Core Design Contradiction:
Object-affected harmful factorsVSLoss of time

Solution Approach 1:

The patent implements a staged training approach where the iterative training process is divided into multiple phases with different levels of iteration depth. In early stages, fewer iterations are performed to establish basic obfuscation capabilities, while later stages perform additional iterations to refine privacy protection. This partial action approach achieves adequate obfuscation quality without the excessive training time required for exhaustive iterative optimization throughout all stages.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3929819B1Method for training obfuscation network which conceals original data to be used for machine learning and training surrogate network which uses obfuscated data generated by obfuscation network and method for testing trained obfuscation network and learning device using the same
Publication Date: 2026.01.28 DEEPING SOURCE INC
  • EP3929819B1 patent drawingFigure 1
  • EP3929819B1 patent drawingFigure 2
  • EP3929819B1 patent drawingFigure 3

AI summary

A method for training an obfuscation network and a surrogate network is provided. The method includes steps of: a 1-st learning device (a) inputting original data of a 1-st party, corresponding thereto, into the obfuscation network to generate obfuscated data wherein the 1-st party owns the original data or is an entity to whom the original data is relegated; (b) transmitting the obfuscated data and the ground truth to a 2-nd learning device corresponding to a 2-nd party, and instructing the 2-nd learning device to (i) input the obfuscated data into the surrogate network to generate characteristic information, (ii) calculate 1-st losses using the ground truth and one of the characteristic information and task specific outputs, and (iii) train the surrogate network minimizing the 1-st losses, and transmit the 1-st losses to the 1-st learning device; and (c) training the obfuscation network minimizing the 1-st losses and maximizing 2-nd losses.