Holistic Moving Target Defence for Autonomous Driving (stage 1a)

Avatar for Xin LOU
Xin LOU    
Associate Professor

Read More 

Avatar for Wei ZHANG
Wei ZHANG    
Associate Professor

Read More 

Avatar for Indriyati ATMOSUKARTO
Indriyati ATMOSUKARTO    
Associate Professor

Read More 

Avatar for Alexander MATYASKO
Alexander MATYASKO    
Researcher
Avatar for Yanghui MO
Yanghui MO    
Researcher

In this project, moving target defence (MTD)-hardened perception systems for object detection (including both regression and recognition), stereo depth estimation, and semantic segmentation will be developed.

Hypernetworks and Monte Carlo dropout are the main approaches considered for implementing MTD. Both real-world physical-space and simulation-based attacks will be used to evaluate the robustness of the MTD-hardened systems. The project will deliver MTD-hardened perception systems, the associated test databases of adversarial examples, real-time demos that contrast unhardened and MTD-hardened systems, and research publications.

Project Deliverables/Outcomes/Impact: 
  • The proposed defence solution won the Best Paper Award in the IEEE Conference on Artificial Intelligence 2025
     
Four-panel comparison of PSPNet, DDC-AT, and ATDE-MTD semantic segmentation models on a road scene with color-coded object masks.
Diagram of a machine learning workflow with two phases: Child Model Diversification (including generation, retraining, and adversarial training) and Prediction Scheduling for pixel-level probabilities.