Synthetic Video Data for Privacy-Preserving ML Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for training machine learning models to monitor individuals using video data often compromise privacy, as manual processing of real video data is required, raising concerns about identity revelation and consent.

Innovation Solution

A method that combines privacy-preserving data feeds from devices like infrared cameras or radars with fake video data generated using generative adversarial networks, ensuring that the training data for machine learning models does not reveal identifiable information, by integrating computer-generated fictive faces into the video data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If real video data is used for training machine learning models, then model training effectiveness is improved, but privacy of the monitored person is compromised

Engineering Contradiction:
Improvemodel training effectivenessVSAvoidprivacy compromise
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent creates synthetic video data that copies the essential characteristics and patterns of real video data without using actual personal images. Generative adversarial networks generate realistic-looking video sequences with synthetic faces that preserve movement patterns, postures, and behaviors needed for training ML models while completely eliminating identifiable personal information.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent introduces synthetic video data as an intermediary between real video data and the ML training process. This intermediary layer allows the model to learn from realistic video patterns without direct exposure to actual personal data, thus mediating between training effectiveness and privacy protection.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Object-affected harmful factors

If privacy-preserving measures are applied to video data, then privacy is protected, but training data quality may deteriorate

Engineering Contradiction:
Improveprivacy protectionVSAvoidtraining data quality
Core Design Contradiction:
Object-affected harmful factorsVSReliability

Solution Approach 1:

The patent changes key parameters of the video data by replacing identifiable features (faces) with synthetic generated faces while preserving critical training parameters such as body movements, postures, positions, and behavioral patterns. This allows privacy protection through parameter modification without deteriorating training data quality.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent applies different quality requirements to different parts of the video data. Identifiable features like faces are completely synthesized with high realism, while critical training features like body movements and postures are preserved with high fidelity from real data, ensuring both privacy and training effectiveness.

Inventive Principle:
Principle #3Local quality

3Measurement precision

If manual processing of video data is performed for training, then training accuracy is improved, but privacy risks increase

Engineering Contradiction:
Improvetraining accuracyVSAvoidprivacy risks
Core Design Contradiction:
Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent employs automated generative adversarial networks that self-generate synthetic video data without requiring manual processing of real personal video data. The GAN system automatically creates realistic training data with proper privacy protection, eliminating the need for human reviewers to manually process sensitive video content.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent replaces the mechanical process of manual video data processing with an automated computational system. Generative adversarial networks automatically generate synthetic training data, substituting human manual review and processing with an algorithmic system that inherently protects privacy while maintaining training accuracy.

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

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

This approach provides valuable training data that maintains the privacy of the monitored individual while effectively training machine learning models to infer states, such as posture and position, without revealing identifiable features, thus addressing privacy concerns and improving model training efficacy.

Implementation Method 1

generating fake video data based on a generative adversarial network, GAN

Methodology Applied
Scientific EffectGenerative adversarial network (GAN):

Data Source

PatentEP4256453B1Providing training data for a machine learning model for monitoring a person based on video data
Publication Date: 2024.07.17 ASSA ABLOY AB
  • EP4256453B1 patent drawingFigure 1~2C
  • EP4256453B1 patent drawingFigure 3~5

AI summary

It is provided a method for providing training data for a machine learning model for monitoring a person based on video data. The method is performed by a training data provider (1). The method comprises: obtaining (40) a data feed of the person, wherein the data feed comprises a series of images that preserves a privacy of the person; generating (42) fake video data of a fictive person, such that a face of the fake video data is a computer-generated face; combining (44) the data feed with the fake video data, resulting in training data; and providing (46) the training data for training the machine learning model.