Abstract
In recent years, applications in the cyber physical systems (CPS) area have greatly benefited from deep learning (DL)'s success. However, there exists an intrinsic problem with directly applying a trained deep learning model on a CPS application: CPS applications have constraints arising from realisms, whereas the training of deep learning models often does not take any or enough realisms into consideration. In the context of this thesis, a realism is defined as the reality as a result of the DL models' interaction with the CPS. Among the multitude of realisms, there are the following types of realisms that are most important. The first realism is task-specific, resulting from the interaction between the deep learning model and the environment in which the deep learning model is deployed. In this thesis, we address this type of realism in a specific acoustic application where audio samples are environmentally distorted in real cyber physical systems (smart homes). The second realism is non-targeted samples, which are defined as samples whose classes are not seen by the DL models during their training. In the same acoustic application, we incorporate a Mahalanobis distance-based out-of-distribution (OOD) detection technique to prevent OOD audio samples from being passed to the classifier trained on in-distribution data. As a result, the classifier is less prone to make mistakes as fewer OOD samples are passed to it. The third realism is that many data-driven deep learning models are not robust against even minor changes. In this thesis, we use an attention-enhanced graph neural network (GNN) architecture coupled with real-world knowledge, using both the GNN architecture and the real-world knowledge as ways to boost the robustness of the DL models. Importantly, the underlying problem with the aforementioned realisms is the problem of domain adaptation (DA): how do we make sure the DL models trained in one domain (clean samples free of these realisms) can perform adequately on another domain (samples tainted by realisms)? In this thesis, we develop two novel unsupervised domain adaptation (UDA) algorithms that are superior to the state-of-the-art UDA algorithms. The final realism addressed in this thesis has nothing to do with the training of the DL models. Instead, it is the realism that comes during the process when the deep learning model is deployed, such as caused by human behavior. In this thesis, we provide a comprehensive analysis of the case study in which we deploy several acoustics-based deep learning models in six smart homes, where we present and evaluate various techniques for deploying the deep learning models in CPS.
