Proj CDeepFuzz Paper Reading: Deepbillboard: Systematic physical-world testing of autonomous driving systems
Abstract
背景:目前的研究集中在例如交换几个像素这种digital perturbation,这些是很难发生在现实世界的;生成both digital and physical adversarial perturbation很重要
本文:DeepBillboard
Task: Test drive-by billboards
Method:
- dynamic changing driving conditions including angle, distance and lighting.
- objective: max the possibility, degree, and duration of the steering-angle errors
实验:
实验1: digital perturbations
效果:DeepBillboard is effective for various steering models and scenes
实验2: physical-world case studies
效果:DeepBillboard is sufficiently robust and resilient for generating physical-world adversarial billboard tests under various weather conditions; being able to mislead the average steering angle error up to 26.44 degrees.