Skip to content Skip to main navigation Report an accessibility issue

Real-time GAN-based Sensor Anomaly Detection and Recovery in Autonomous Driving

Shahrbanoo Rezaei
ISE 3rd Year PhD student
University of Tennessee-Knoxville
Friday, October 15, 2021  3:30-4:30pm


Autonomous vehicles are an integral part of the future of the intelligent transportation systems. The safe operation of these vehicles highly depends on the data they receive from their external or on-board sensors. Autonomous vehicles like other cyber-physical systems are subject to cyberattacks and may be affected by faulty sensors. The consequent anomalous data can risk the safe operation of autonomous vehicles and may even lead to fatal accidents. Hence, it is imperative to design a robust real-time anomaly detection and recovery approach that can mitigate various types of anomalies that may arise as the vehicle navigates the dynamic road environment. To this purpose, we develop an unsupervised adversarial machine learning approach to address this gap. Specifically, we develop an approach based on generative adversarial networks (GANs) that uses a generator to learn the distribution of non-anomalous data to detect and recover different types of anomalies in autonomous driving. This GAN-based approach is evaluated using the Lyft Level 5 dataset. The numerical experiments illustrate that the proposed approach outperforms state-of-the-art benchmarks, regardless of the anomaly types, anomaly rates, and sensors/data streams affected.


Shahrbanoo Rezaei is a third-year Ph.D. student in the Department of Industrial and System Engineering at the University of Tennessee, Knoxville. She works under Dr. Anahita Khojandi supervision and is scheduled to graduate in summer 2023. Rezaei’s research focuses on improving the application of deep learning and reinforcement learning to the safe trajectory of autonomous vehicles. She specialized in solving anomaly detection problems in autonomous vehicles through the use of machine learning models, especially generative adversarial networks, which is a generative model that can generate synthetic but realistic images. She has also worked on the motion prediction and safe trajectory planning in autonomous vehicles using deep learning methods such as ResNets and recurrent neural networks. Although her research mainly focuses on the safe operation of autonomous vehicles, she has also worked on the improvement of park-and-ride facility and services in metropolitan areas of Tennessee supported by Tennessee Department of Transportation (TDOT). She also works together with Oak Ridge National Laboratory on the physics-guided machine learning models for improving aerospace manufacturing, which is supported by DOE.