Master thesis presentation Hannes Johansson & Daniel Olander, MPCAS

Abstract: Driver Assistance Systems (ADAS) and Autonomous Driving Systems (ADS) increasingly rely on Deep Learning (DL) models. While DL models achieved state-of-the-art performance for a variety of tasks, they are not robust across a wide range of traffic scenarios, require large quantities of continuously collected data, and must follow data privacy regulations. Federated Learning (FL) can be used to harness more data from modern vehicles equipped with visual sensors. FL enables training to be done locally at each client (vehicle), sharing and aggregating the trained models while keeping the data local. It also enables models to be trained across regions, limiting sensitive data sharing while training on diverse datasets. This master thesis evaluates the performance of multiple FL algorithms for trajectory prediction (a central part of ADAS/ADS) compared to a central learning approach. A FL framework was implemented and validated through classification experiments on the MNIST dataset using a convolutional neural network. The performance of a central model (one client) was used as a benchmark, achieving a validation accuracy of 99 %. Results show that FL with increasing numbers of clients takes longer to converge but eventually saturates at a similar level of performance. Independent and identically distributed (IID) data yielded the best performance, while non-IID data introduced more noise and overall lower performance. Decreasing the client fraction, the fraction of selected clients for each round of FL, to less than 1.0 corresponded to decreasing performance in the case of non-IID data, but increased training speed. To test the effect of FL algorithms on trajectory prediction performance, the Nuscenes dataset, a collection of visual data from vehicles driving in Boston and Singapore, was used. The data was transformed into 2-dimensional images and fed to CoverNet, a model based on the residual network ResNet-50. The results from the trajectory prediction experiments (varying federated optimization algorithms [FedAvg, FedAvgM, FedProx], client fraction, number of clients and IIDness) followed the trend of more clients resulting in slower training, but overall similar maximum performance. Reducing the client fraction resulted in improved training speed and, in some cases, performance. Non-IID data, i.e., sharing the respective data of Boston and Singapore to different clients, did not result in decreased performance.

Password: 437887

Supervisor: Koen Vellenga
Examiner: Mats Granath
Opponent: Alicia Rey Alonso