With the prevalence of Internet-of-Things (IoT) devices, IoT sensing technologies are also quickly evolving and gradually transforming our lifestyles. Deep Learning techniques, which have shown great success in many areas, are also being adopted in IoT sensing applications. However, we may not have enough data for Deep Learning model training, or only have unlabeled sensor data, which is very common in IoT sensing scenarios. As an important branch in Transfer Learning, Domain Adaptation and Generalization techniques can be applied in such scenarios to train the target model with the help of related source datasets. However, currently there is not much work that explores the application of Domain Adaptation and Generalization in IoT sensing scenarios. Their effect on real IoT sen...[ Read more ]

With the prevalence of Internet-of-Things (IoT) devices, IoT sensing technologies are also quickly evolving and gradually transforming our lifestyles. Deep Learning techniques, which have shown great success in many areas, are also being adopted in IoT sensing applications. However, we may not have enough data for Deep Learning model training, or only have unlabeled sensor data, which is very common in IoT sensing scenarios. As an important branch in Transfer Learning, Domain Adaptation and Generalization techniques can be applied in such scenarios to train the target model with the help of related source datasets. However, currently there is not much work that explores the application of Domain Adaptation and Generalization in IoT sensing scenarios. Their effect on real IoT sensing problems is not fully studied. Moreover, current works do not consider the model training problem in real IoT environment. With the increased attention of the public on data privacy, it becomes more and more difficult to collect all data in a data center for centralized training. Therefore, a distributed model training framework is needed for practical Domain Adaptation and Generalization in IoT environment.

In this thesis, we first describe our exploration to apply Domain Generalization to a practical IoT sensing problem, which is driver monitoring using FMCW radar. We tested 2 different Domain Generalization methods, namely CCSA and MMD-AAE, and their modifications, in an attempt to reduce the amount of real driving data needed and make the model perform well on unseen drivers and cars. Then we will introduce our design of a Federated Domain Adaptation framework, which is a distributed training framework tailored for Domain Adaptation training in IoT environment. It incorporates Federated Learning and Homomorphic Encryption to protect data privacy of all participants. Moreover, we design 2 optimization methods which further reduce the computation and communication overhead during training, making our framework more friendly to resource-constrained IoT devices.