User-Behavior-aware Real-Time Safety Enhancement in Vehicular Networks

Principal Investigator: 
Sai Manoj Pudukotai Dinakarrao, assistant professor, electrical and computer engineering, George Mason University

Co-Principal Investigators: 
Haiying (Helen) Shen, associate professor, computer science, University of Virginia

User-Behavior-aware Real-Time Safety Enhancement in Vehicular Networks Project Description:

The Machine Learning (ML) models in autonomous vehicles and vehicular networks need to be adaptable, secure and robust for the purpose of safety and security of autonomous driving. One-shot learning has recently been introduced to enhance the adaptability of autonomous vehicles in different environments through the generation of synthetic data. However, due to a lack of environmental feedback to tune its hyper-parameters, there is no guarantee that this synthetic is accurate or comprehensive enough to cover all conditions to which an ML model in an autonomous vehicle may be exposed.

The CCI research team is addressing this challenge through the development of a novel online one-shot learning that considers user-behaviors as feedback for better adaptability and robust learning, focusing on three areas of advancement to this field. First, researchers will develop a one-shot learning algorithm that generates synthetic data based on direct feedback from user behaviors, which previous one-shot online learning algorithms do not. Second, as a part of this algorithm, researchers will deploy an adversarial robustness mechanism that utilizes user behaviors for model training, further enhancing the robustness and security of the underlying ML models to adversarial attacks. Finally, researchers will identify and implement defense techniques against data stealing attacks to secure and enhance the ML models. The application of this research could contribute to the emerging autonomous vehicular field – specifically the development of autonomous driving and vehicular networks – and showcase the importance of considering user behavior during the learning of ML models.