| Summary: | Autonomous vehicles are becoming increasingly popular due to the potential of eliminating vehicular accidents stemmed from human error. The computers that control the vehicles make purely logical choices based on data gathered, in contrast to human judgement which might be affected by poor reaction, distractions, or substances such as drugs or alcohol.
Autonomous vehicles use computer vision to gather information and navigate the environment. Data gathered will come from in the form of images of its environment, including traffic lights, pedestrian, stop signs etc. These images are processed through a deep neural network, which makes decisions based on the results it churns from the data gathered. As these vehicles rely heavily on these images to determine the situation on the road, it is important the data retrieved needs to be accurate.
However, recent research shows that adversarial examples, which are small perturbations added to real images can mislead deep neural network-based computer vision algorithm. This poses a problem as malicious parties could mislead autonomous vehicles into making wrong decisions.
Hence, this project will construct an autonomous vehicle testbed to facilitate the research of adversarial examples that is used in development of countermeasures. The testbed consists of LGSVL simulator and Apollo driving agent. This project attempts to retrieve the model from Apollo and test it against adversarial examples. The greater the effectiveness of these adversaries, the more relevant the research is towards the development of countermeasures.
|
|---|