Towards Secure Deep Neural Networks for Cyber-Physical Systems

In recent years, deep neural networks (DNNs) are increasingly investigated in the literature to be employed in cyber-physical systems (CPSs). DNNs own inherent advantages in complex pattern identifying and achieve state-of-the-art performances in many important CPS applications. However, DNN-based systems usually require large datasets for model training, which introduces new data management issues. Meanwhile, research in the computer vision domain demonstrated that the DNNs are highly vulnerable to adversarial examples. Therefore, the security risks of employing DNNs in CPSs applications are of concern.

In this dissertation, we study the security of employing DNNs in CPSs from both the data domain and learning domain. For the data domain, we study the data privacy issues of outsourcing the CPS data to cloud service providers (CSP). We design a space-efficient searchable symmetric encryption scheme that allows the user to query keywords over the encrypted CPS data that is stored in the cloud. After that, we study the security risks that adversarial machine learning (AML) can bring to the CPSs. Based on the attacker properties, we further separate AML in CPS into the customer domain and control domain. We analyze the DNN-based energy theft detection in advanced meter infrastructure as an example for customer domain attacks. The adversarial attacks to control domain CPS applications are more challenging and stringent. We then propose ConAML, a general AML framework that enables the attacker to generate adversarial examples under practical constraints. We evaluate the framework with three CPS applications in transportation systems, power grids, and water systems.

To mitigate the threat of adversarial attacks, more robust DNNs are required for critical CPSs. We summarize the defense requirements for CPS applications and evaluate several typical defense mechanisms. For control domain adversarial attacks, we demonstrate that defensive methods like adversarial detection are not capable due to the practical attack requirements. We propose a random padding framework that can significantly increase the DNN robustness under adversarial attacks. The evaluation results show that our padding framework can reduce the effectiveness of adversarial examples in both customer domain and control domain applications.