Zenseact and Syntonym Use Generative AI to Improve Autonomous Driving Training Data Privacy

On June 16, Zenseact and Syntonym announced a new approach that uses generative AI to replace faces in autonomous driving training data instead of blurring them. The initiative is designed to protect privacy while retaining the visual information required for advanced vehicle perception systems. By maintaining critical behavioral cues, the solution aims to improve the quality of training data used by autonomous vehicle developers without exposing personally identifiable information.

Balancing Privacy and Data Quality in Autonomous Driving

Autonomous vehicles depend on extensive video datasets to learn how pedestrians, cyclists, and other road users behave in real-world environments. These datasets are essential for training perception models that must accurately interpret human actions and intentions. However, privacy regulations and ethical considerations require that identifiable information, including faces and license plates, be protected before the data can be used for development and testing purposes.

Traditional anonymisation methods typically rely on blurring sensitive areas within images and videos. While effective for privacy protection, blurring often removes valuable information such as eye direction, facial movement, posture interactions, and expressions. These subtle indicators help AI systems understand intent, predict movement, and make safer driving decisions. As a result, developers face a trade-off between privacy compliance and dataset utility.

Generative AI-Based Synthetic Face Replacement

Syntonym's technology introduces a different approach by generating synthetic facial replacements for real individuals captured in training footage. Instead of obscuring faces, the system creates artificial substitutes that cannot be linked back to the original person. At the same time, important physical and behavioral characteristics are preserved, allowing machine learning models to continue learning from realistic human interactions.

The synthetic faces maintain visual consistency within the dataset while eliminating the risk of identifying individuals. This enables autonomous driving systems to access richer information during training, potentially improving their ability to recognize intent and respond appropriately to dynamic traffic scenarios involving pedestrians and cyclists.

Enhancing the Zenseact Open Dataset

The immediate objective of the collaboration is to strengthen the Zenseact Open Dataset, a large multi-modal dataset used for autonomous driving research and development. By incorporating synthetic anonymisation techniques, the dataset is expected to provide higher utility than traditional privacy-preserving approaches while continuing to meet privacy requirements.

According to the companies, the adoption of synthetic anonymisation has the potential to move beyond current industry standards for balancing privacy and data usefulness. The initiative demonstrates how generative AI can support both responsible data management and the development of more capable autonomous driving systems.

Frequently Asked Questions

Why are faces usually anonymised in autonomous driving datasets?
Faces are anonymised to protect personal privacy and comply with data protection requirements when collecting real-world driving footage. Autonomous vehicle developers gather large amounts of video data containing pedestrians, cyclists, and other road users. Without anonymisation, individuals could potentially be identified from the recordings. Privacy-preserving methods ensure that training datasets can be used for AI development while reducing risks related to personal data exposure and regulatory compliance.

How does synthetic anonymisation differ from traditional blurring?
Synthetic anonymisation replaces real faces with AI-generated alternatives rather than obscuring them with blur effects. This approach protects identity while preserving visual details such as gaze direction, facial movement, and expressions. These characteristics are important for helping autonomous driving systems understand human intent and behavior. As a result, developers can maintain higher-quality training data while still meeting privacy objectives and safeguarding individuals captured within the dataset.