Ignore parts of image during CNN training

Question

I have a traffic dataset (similar to KITTI) where license plates of cars are masked with white boxes due to anonymization purposes. Labels of vehicles (Bounding Box + Type) are