Impact of Cyber Attacks on Traffic State Estimation for Connected and Autonomous Vehicles (CAVs) Systems
Primary Investigator:
Satish Ukkusuri
Eunhan Ka
Abstract
As technology continues to evolve, we will see a growing number of connected and autonomous vehicles (CAVs) on road networks. CAVs can improve traffic operations, revolutionize transportation systems and reduce road accidents. Network traffic flow models have a significant role in enhancing the efficacy of traffic management strategies by estimating traffic states and describing traffic dynamics. Despite having robust theoretical foundations, existing network traffic flow models struggle to model the complex and dynamic real-world traffic data - especially the variance and heterogeneity in large-scale urban networks. The physics-informed deep learning model with a generalized bathtub model (PIDL-GBM) leverages the interpretability of physical models and ML methods for their powerful modeling ability. The input data of PIDL-GBM is crucial to learn the invisible relationship between input and target variables in the training process. However, CAVs systems pose a significant cybersecurity risk. Escape attacks, one of the attacks on machine learning systems are craft input data of machine learning models (e.g., removal, manipulation). This study aims to quantify the impacts of cyber attacks on traffic state estimation with PIDL models. We test the proposed method on mobile location data and a large-scale road network in Indianapolis, United States, with various attacks ratio. The experimental results show that escape attacks significantly deteriorated the performance of the PIDL-GBM model as the attack ratio increased.