Get Demo
Use cases
Industry

Industry

PLAY PAUSE
0:00
/
PLAY PAUSE

Privacy-Preserving Machine Learning for Industries

Industry Domains:

Transportation, Mining and Metallurgy

Encryption method:

Functional Transformation (Veils)

Data type:

image, video

ML-model:

YOLOv7, YOLOv8

ML-tasks:

Computer Vision for detection of:

  • the fact, sequence, and duration of tool usage in a production chain;
  • the use of helmets, safety glasses, gloves, and protective footwear for compliance with safety regulations.
Challenges:
  • Data protection
  • Maintaining quality
  • Inference protection
  • Model protection

Customers

Three major companies from the Transportation, Mining, and Metallurgy sectors. Large infrastructure firms manage a vast network of buildings, facilities, and structures distributed across the globe.

Their employees perform daily tasks related to the technical maintenance of these facilities and the communication systems connecting them.

The companies' objective is to monitor the execution and quality of these tasks, as well as ensure compliance with safety regulations. The geographic dispersion of the facilities further complicates this task.

Among the common attributes of Сustomers are:

  • Regularly engaging in R&D and the implementation 
of machine learning

  • Exploring the use of third-party cloud platforms for these purposes

  • Intending to involve external teams of ML specialists

  • Relying on their own remote specialists

Challenge

Data related to the use of tools in the production chain and compliance with safety regulations is classified by Customers as sensitive. This is because the video footage captures employees, facilities, and objects, as well as incidents that require thorough investigation.

To efficiently work with ML algorithms, it is essential to involve external specialists, utilize powerful computational resources offered by cloud providers, and grant access to data for employees working outside the clients' security perimeter.

For this, it is necessary to:
  • Protect data during ML model training on third-party cloud platforms, when working with outsourced ML teams, and with the internal remote ML team,
  • Maintain a quality level comparable to training on open data,
  • Protect inference,
  • Safeguard the model as a result of intellectual work.

Functional Transformation (Veils)

The data owner transforms it into a non-interpretable form using a one-way transformation function

Transfer  of non-

interpretable data to

the ML team

Storage of non-

interpretable data

Training  an ML model

on non-interpretable

data

ML-model

quality check

Protection of data

sent to the inference

Solution

To address the outlined challenges, the Veils method, which belongs to the class of functional transformation methods, was applied.

For the purpose of publicly demonstrating the aggregated results of three different clients, we used the open dataset "Mechanical tools-10000" with five detectable classes. ML experiment tracking was conducted in TensorBoard.

The graphs below display the mAP_0.5 metric during the training of the YOLOv7 model. The pink graph represents training on the original dataset images, while the violet graph corresponds to training on data protected using the Veils method.

As shown in the graphs, the difference in the mAP_0.5 metric does not exceed two percent.

The graphs for other metrics also show similar behavior.

01 \ 09
The following table provides some parameters and characteristics of the training process.
Input data Batch size Epochs Workers Max mAP_0.5 Max
mAP_0.5:0.95		 Run time
(hours)		 Dataset size
Original
images		 10 100 8 0,9082 0,7307 6 332 MB
Veils 10 100 8 0,8917 0,717 7,2 53000 MB

The computation time of protected data for the entire dataset was 20 minutes on a PC with a GPU and 70 minutes on a PC without a GPU.

 

Below are the results of applying the trained models to the test dataset of images.

Based on the values provided in the tables, it can be concluded that the Veils data protection method is industrially applicable.
Results obtained on the original images
Results obtained on data protected using the Veils method

Results

During training, the difference in mAP_0.5 does not exceed 2%.

Results on the test data show that the Veils method, which belongs to the class of functional transformation methods, can be industrially applicable.

The outcome of the projects for clients was the development of ML models trained on protected data for detecting required objects in images.

The solution ensures the subsequent safe transfer of confidential data. 
This data, collected from numerous distributed video recording points of Customers employees' actions, is sent to local data centers where protected inference is performed on the trained ML model.

 

This, in turn, allows for monitoring the fact and duration of infrastructure maintenance tasks, as well as compliance with safety regulations.

 

It also reduces costs for monitoring line personnel, optimizes production processes, and decreases accidents and workplace injuries.

Clients received quantitative data for a more justified employee motivation and reward system and for more favorable negotiations with insurance companies.

The models, as a result of intellectual work, have been protected.

Please review the additional materials that support the value of integrating computer vision 
and machine learning technologies in the industry:

[1] Computer Vision for Safety Management in the Steel Industry
[2] Applications of Machine Learning and Computer Vision in Industry 4.0
[3] Role of Computer Vision in Industry 4.0: Top 5 Use Cases Transforming Manufacturing Industry
[4] Complete Guide to AI-Powered Workplace Safety in Manufacturing
[5] Computer Vision in Workplace Safety
[6] Role of Computer Vision (CV) in Manufacturing
[7] Leveraging computer vision towards high-efficiency autonomous industrial facilities

Give it a try yourself!
Request a Demo Version of the Data Protection Software here