Towards optimal defences on adversarial examples for DNN-driven digital twinning

• Main Author: Lee, Michael Yew Chuan
• Other Authors: Jun Zhao
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/171756
• Institution: Nanyang Technological University

Digital twinning is one of the main enablers of the Metaverse. It involves the creation of a digital twin (DT), a virtual model that accurately reflects a physical entity (PE) in real time. Integral to digital twinning are DNNs, which play a pivotal role in enhancing the digital twinning process. Not only are DNNs used to fulfil the functional requirements of DTs, but they also facilitate essential underlying processes supporting DTs. This includes enabling seamless information communication and optimising the allocation of resources among the devices that support DTs. Thus, DNNs are crucial to the optimal and smooth execution of DTs. However, DNNs are vulnerable to a type of attack known as adversarial examples. Such attacks threaten the functionality of DTs when DNNs supporting the digital twinning process are attacked. While defences for DNNs exist, works typically only focus on high attack prevention rates. However, tradeoffs exist when applying these defences in the real world. While high attack prevention rates lessen the threat to DTs, it could lead to increased network latency and resource usage. These effects if significant can negatively impact the functionality of DTs, and harm the digital twining experience. As such, we argue that it is equally important to consider the tradeoffs when applying defences in the real world. This will ensure the real-time support required by DTs in the Metaverse. In this paper, we begin by discussing adversarial attacks and defences. Then, we show how the entire DNN-enabled digital twinning pipeline is susceptible to attacks, and suggest defences to defend against them. Following this, we introduce a framework that uses deep reinforcement learning as an optimiser to alleviate the tradeoffs that arise from implementing the defence mechanisms. This will improve the feasibility of defences for DNNs supporting the digital twinning process. Experiments demonstrate that our solution can alleviate the tradeoffs incurred.