University: Chalmers University of Technology

Data collected with the sensors of any car on the road could be vital to realize advanced safety innovations. Currently, it is not possible to learn from all processed or collected data by cars on the road. Meanwhile, the robustness of advanced safety features could benefit from learning from more collected data including unlikely scenarios. The main obstacles to using all the data are the costs of the data transfer to a central server and data privacy regulations. Therefore, the purpose of this thesis project is to explore the influence of federated deep learning approaches to mitigate this, in collaboration with Volvo Cars.

Modern cars use multiple sensors (e.g., radar, lidar, cameras) to collect large amounts of data, and are equipped with specialized hardware to accelerate the data processing. Although researchers and companies have collected more and more data to improve safety and autonomous driving functionalities, there is still a large amount of data that is processed by cars on the road not being used for improving car safety functionalities. The main obstacles to use data from all cars on the road are the costs of the data transfer to a central server and data privacy regulations. One of the major unsolved problems in machine learning (ML) safety is the research in long-tail robustness to also support unlikely scenarios. 

To surpass the need to transfer and store data, one could train ML models locally in a car and only transfer the trained model weights. The federated learning (FL) concept is a different strategy to train ML models and introduced. Traditionally one would use all the labeled data on a central computer or server and run the training algorithm and distributes the centrally trained model. In a FL setting the ML models are trained on the device that collects the data (in our scenario within the car), which removes the burden to transfer and store data, and only requires one to transfer the locally trained ML model. The locally trained models can be aggregated via different strategies on a central server and re-distributed to the devices. 

We acknowledge that multiple challenges must be overcome before FL can be adapted into a wider automotive setting. These challenges are, for example, potential privacy leakage, large variance in the data that which prohibits a model to convergence, more data does not directly lead to increased performance, and automatic labelling can cause misclassification or introduce biases. While not necessarily solving all these issues, we focus on empirically establishing if FL has potential in an automotive setting,  contributing to the understanding whether FL trained models should be a considered approach for automotive safety techniques.