Tiresias

Police

This algorithm shields personal data in police data by automatically blinding vehicle windscreens. This is done when vehicle images are requested as part of an ongoing investigation.

Last change on 31st of January 2025, at 13:04 (CET) | Publication Standard 1.0

• Publication category: Other algorithms
• Impact assessment: DPIA
• Status: In use

General information

Theme

Public Order and Safety

Begin date

2022-10

Contact information

https://www.politie.nl/

Responsible use

Goal and impact

The aim of Tiresias is to protect the privacy of citizens. This is done by blinding the windscreen of a vehicle, making the occupants unrecognisable.

Considerations

The police themselves took the initiative to automate the manual blinding of windscreens. This process was physically stressful for employees, time-consuming and proved less reliable on inspection. Automatic shielding saves time and ensures consistency. Since vehicle images are usually taken from the same angle, an algorithm can perform it efficiently and accurately.

Human intervention

The collaborator checks the work of the algorithm. All images are shown to the worker. If necessary, the worker can blind additional parts of the image, but not undo what has already been blinded.

Risk management

The entire windscreen is blinded instead of just the people behind it. This is easier and ensures that nothing behind the windscreen is visible, such as faces, clothing or other details. Only the vehicle and the registration number remain visible.

Legal basis

In the Code of Criminal Procedure (Section 126jj), the police are empowered to record the registration number, time, location and photograph of a vehicle on the public highway in certain cases and under specified circumstances.

Links to legal bases

• Wetboek van Strafvordering - Titel VF. Vastleggen en bewaren van kentekengegevens - Artikel 126jj.: https://wetten.overheid.nl/BWBR0001903/2025-01-01#BoekEerste_TiteldeelVF
• Besluit vaststelling nadere regels vastleggen en bewaren kentekengegevens ex artikel 126jj Wetboek van Strafvordering door politie.: https://wetten.overheid.nl/BWBR0041691/2019-01-01

Impact assessment

Data Protection Impact Assessment (DPIA)

Operations

Data

The Tiresias algorithm is trained on images of vehicles, also known as ANPR (Automatic Number Plate Recognition) images. These images show the vehicle and its registration number, and sometimes people behind the windscreen. This depends on factors such as the time of day and light.

Technical design

A machine learning model determines which parts of the image should be blinded. This is done with a special algorithm (a neural network) that creates a 'mask' to mark the areas to be blinded (semantic segmentation). A UNET model [1] is used for this purpose, which is trained to accurately recognise areas such as vehicle windscreens. During training, the algorithm itself learns how to mark windscreens.

The image is pre-adjusted to a size of 224x224 pixels (configurable) and the brightness is normalised. As a result, the model produces an image in which all vehicle windscreens are marked.

After the algorithm is trained and it is applied, the algorithm does not continue to learn. It will then always give the same output at the same input.

Notes for experts:

* A UNET model with Resnet18 backbone is used. [2]

* The model is trained with a combination of Binary Cross-Entropy (BCE) and Jaccard-loss.[3]

Sources:

[1] Ronneberger O, Fischer P, Brox T (2015). "U-Net: Convolutional Networks for Biomedical Image Segmentation".

[2] He, Kaiming; Zhang, Xiangyu; Ren, Shaoqing; Sun, Jian (2016). Deep Residual Learning for Image Recognition. Conference on Computer Vision and Pattern Recognition.

[3] MA Rahman and Y Wang (2016). Optimizing intersection-over-union in deep neural networks for image segmentation. International Symposium on Visual Computing.

External provider

The neural network is fully trained on its own data and does not use pre-trained open-source models. The source code used to program this network comes from an open-source project.