Continuous, active Federated Learning for data streams

This project aims to advance safe automated vehicles by enhancing AI-driven perception, situational awareness, and decision-making while mitigating privacy breaches and national security risks.

Vasagatan street in Stockholm with pedestrians, cars, and bicyclists.

Traffic in Stockholm, Photo by Agata Ciosek on Unsplash

Introduction

The increasing use of edge devices in vehicles (sensors such as LiDAR, cameras, etc.) has led to massive data generation, creating challenges for real-time AI adaptation, data privacy, and security.This project targets federated learning for real-world data streams at the edge, addressing challenges related to both continuous and active learning methods in practice.

Challenges

Existing FL frameworks have primarily been designed for static datasets, requiring devices to store and iterate over local data before updating global models. However, this assumption does not hold in automotive applications, where:

Vehicles continuously generate sensor data at high frequencies.
On-board storage and computational power are limited.
Real-time adaptation is critical for autonomous driving and Advanced Driver Assistance Systems (ADAS).
The storage and transmission of sensor data from vehicles introduces significant security risks, particularly in cases where vehicles capture sensitive locations (e.g., government buildings, military installations, private properties).

This project will address these domain-level challenges by developing learning methods that work under strict, real world conditions. This presents the following technical challenges:

Designing active learning components that work over small scale data streams.
Integration of those active learning components in a federated learning system - making models learn on extremely low context (data streams rather than data sets).
Assessing catastrophic forgetting and handling data drift—what happens when small edge models start to forget existing learnings to accommodate new ones?
Leveraging low-context learning to enhance the security of the learning mechanism.

As the automotive industry continues work on AI-driven automation, ensuring that data can be leveraged for safer driving without compromising privacy or security is paramount. This project pioneers a new approach to real-time federated learning, enabling continuous AI adaptation while fully complying with increasing regulations. By tackling these challenges, we are not just advancing autonomous driving—we are shaping the future of safe and secure, intelligent mobility.

Jonas Ekmark

Head of New Technology, Zenseact

Expected outcomes

To advance AI-driven solutions for traffic safety in ADAS and Autonomous Driving (AD), it is crucial to develop data collection and AI training methods that don't compromise privacy and national security, and this is what we expect this project will deliver.

This project will create the necessary solutions to ensure that Swedish companies can continue innovating in AI-driven mobility while fully complying with evolving regulatory standards, protecting both individual and national interests. The resulting technology is not limited to automotive applications and will be disseminated for broader use in Swedish industry.

We believe that such a streaming-based federated learning solution is critical for the next generation of safe, intelligent transportation systems, ensuring that models improve continuously and securely while complying with regulatory constraints.

We expect that the results will inform regulatory discussions on AI-driven traffic safety, supporting the development of policies for secure, automated vehicles.