SysConTalks is an initiative by the Ph.D. students of the Systems and Control group at IIT Bombay. We organize seminars by researchers at the forefront of Academia and the Industry, and our aim is to provide an engaging platform for students such as ourselves to learn from the best minds in the broad areas of systems and control, optimization, and data science.
Never miss an update: Click here to join the google group
Dr. Arpan Kusari
University of Michigan Transportation Research Institute
Title: Generating broken lens Camera failures as a class of physics-based adversarial examples
Schedule: 16 December 2024
Venue: SysCon Seminar Room
Abstract
Although there has been significant research on generating physics-based adversarial samples recently, one often overlooked category of such samples stems from physical failures within the camera itself. In this study, we develop a simulated physical process to create broken lens scenarios, exemplifying a type of physics-based adversarial sample rooted in camera failures. We employ a finite element model (FEM)-based simulation to generate surface cracks by applying a stress field defined by particles constrained in a triangular mesh. With physically-based rendering (PBR), we realistically visualize the resulting glass cracks. To understand the effect of these adversarial samples, we apply real instances of broken lenses and generated adversarial samples as image filters on two different open-source datasets, KITTI and MS-COCO. We measure the detection failure rate of various classes using two object detection neural network models, YOLOv8 and Faster R-CNN. To understand the distributional differences between the images, we compute the Kullback-Leibler (K-L) divergence across three different data distributions using various image filters: a dataset we collected through a windshield crack, KITTI, and images of cats from the Kaggle cats and dogs dataset. The K-L divergence results indicate that the different broken lens filters are not significantly different in terms of distribution. Our objective with this work is to establish a robust physics-based process for generating adversarial samples.
Biography
Dr. Arpan Kusari is a research faculty at University of Michigan Transportation Research Institute (UMTRI) specializing in systemic issues of robustness in autonomous vehicles (AV). In this role, he has been working with federal sponsors such as National Highway Traffic Safety Administration (NHTSA) and Federal Highway Administration (FHWA) and corporate sponsors such as Ford Motor Company, Collaborative Safety Research Center (CSRC) and Subaru Motor Company to name a few. His team works on analyzing robustness issues arising in the sensing, perception and decision-making sub-systems from a regulatory standpoint and providing practical and scalable solutions. Dr. Kusari has been selected as one of the winners of the U.S.D.O.T. Safety Intersection Challenge and the winner of the TRB Data Forecasting Challenge 2022. Previously, at Ford Motor Company, he worked as a research engineer on developing model-based solutions for mapping, localization, behavioral prediction etc.