RT4: AI Driven Edge Data and Model Guard

Theme Leaders: Dr Varun Ojha & Dr Yang Long

Researchers: Prof Dhaval Thakker and Prof Xianghua Xie

RT4 develops and innovates methods, techniques, and algorithms to monitor, recover, and ensure the continuity of high-quality AI data and AI models in deployed environments. RT4 develops algorithms to safeguard AI data and models from cyber disturbances and attacks in edge computing systems. We ensure the developed algorithms are privacy-aware, secure, robust, and resource-efficient when deployed at the edge.

We work with RT2 and RT3 to develop AI models that can defend against both known and unknown cyber threats. This means tackling challenges like monitoring how cyber issues affect data and AI models, recovering from unknown problems, and ensuring our solutions work on all kinds of edge devices.

We create ways to continuously monitor edge computing systems so we can respond quickly to cyber problems in real time. This is tricky because edge systems can be complex and have limited resources. We use what we learn from RT2 to decide what to monitor and ensure our tools can handle different kinds of systems. We develop flexible, community-driven methods to protect data and models from cyber threats. We use advanced techniques like agentic AI, generative AI, and reinforcement learning, which let AI models learn and adapt over time.

We engage with industry and other stakeholders, gather expert feedback, and update our protection methods to keep pace with emerging threats and rapid advancement of AI. 

The Vision of RT4:

The vision of RT4 (AI Data and Model Quality) is to establish research directions for developing fundamental concepts and techniques that can guard the data and AI algorithm learning quality against cyber-disturbances impacting EC architectures.

This theme sets out the following: Challenges / Research Aims:

• Monitoring of Data/Model Quality

How to monitor cyber-disturbances impact on the quality of data, AI algorithms learning and the overall application resilience?

• Recovery of Data/Model Quality

How to ensure the recovery of data and AI model quality that are impacted by cyber-disturbances and ensure suitability for AI model deployment on devices at Tiers 1, 2 of EC architectures?

• Assurance of Continuity of Data Quality and Model Quality

How to assure AI algorithms continually adapt to EC environments where unknown cyber-disturbances that were not present in the original training dataset?

Figure 1: Overview of RT4 (AI Data and Model Quality)

Video 1: Drowned Drone, an example work in RT4, that enable safeguarding AI data and model quality

Figure 2: Data poisoning adversarial attacks on deep learning algorithms (more on https://doi.org/10.1016/j.artint.2023.104060)

Figure 3: Quality of Graph Transformer: An experiment with Low-rank and global-representation-key-based attention for graph transformer (more on https://doi.org/10.1016/j.ins.2023.119108) 

Figure 4: Alignment approach for deep learning model quality improvement (https://doi.org/10.1109/TPAMI.2024.3487631)

Figure 5: Alignment approach for deep learning model quality improvement

Animation 2: pre-and post- buckling analyses of space truss structures in civil engineering using artificial intelligence methods