Byzantine Fault Tolerant Federated Learning for Machine-to-Machine Communication: A Smart City Vehicular Network
Abstract: A common conception of smart vehicular and transportation systems is to create a fault-tolerant pair-to-pair network among the vehicles in the communication range such that they can share information about traffic, road conditions, and many more. Although there are several hurdles to overcome, recently most vehicles are equipped with LIDAR, ultrasonic, and camera sensors for the Advance Driving Assistant System (ADAS). Two major problems in developing a machine-to-machine network with these sensors are the sheer amount of data (gigabytes per second) that needs to be sent over the communication channel and some of the nodes may not respond or can provide inaccurate Data. However, the end goal of this data is to train a certain ML or DL model for prediction and Automation. Thus using federated learning techniques in each vehicle's onboard computation unit, the main model can be trained, and each of the model parameters (A few KB) can be sent over the network instead of the data. Then this newly created model information can be validated using a specific consensus algorithm to remove the BFT problem. Central Cloud storage might also be helpful to accommodate the overall ML/DL model version history such that if any of the models proved to be falsified by misleading training data, a quick rollback can be possible