No abstract

No abstract

We present the first backdoor attack against the lane detection systems in the physical world. Modern autonomous vehicles adopt various deep learning methods to train lane detection models, making it challenging to devise a universal backdoor attack technique. In our solution, (1) we propose a novel semantic trigger design, which leverages the traffic cones with specific poses and locations to activate the backdoor. Such trigger can be easily realized under the physical setting, and looks very natural not to be detected. ( 2) We introduce a new cleanannotation approach to generate poisoned samples. These samples have correct annotations but are still capable of embedding the backdoor to the model. Comprehensive evaluations on public datasets and physical autonomous vehicles demonstrate that our backdoor attack is effective, stealthy and robust.

No abstract

Perception Testing technologies are widely applied in various scenarios, like industrial and academic research applications for autonomous driving systems. Accurate and robust autonomous driving simulation perception is pivotal for safety-guidance autonomous vehicles (AV). The autonomous vehicle system is facing the main challenges of a complex real-world environment with multi sensors' performance and their neighborhood view with an uncertain environment. The perception module exploits deep learning models to detect surrounding obstacles, including their types, positions, and velocities. However, the issue of LiDAR performance, which is the prediction of multi obstacle properties based on their deep learning model, remains unresolved. In addition, autonomous Vehicle systems rely on deep learning models to collect and modify raw point cloud data. It's crucial to check autonomous vehicles' robustness in several scenarios to generate automotive vehicles more safety and reliable. At this stage, it is easier to perform sensor integration to simulate sensors' performance for Autonomous Driving at various levels of verification. This thesis utilizes a realistic LiDAR point cloud and compares the difference between real-world and simulation environments to test perception modules for autonomous driving platform systems. This thesis proposes a simulation-based testing platform for autonomous vehicles to discover the potential shortage of perception by analysing huge scenarios and testing the suitable sensor performance in multi particular levels of the automated driving platforms. Additionally, this thesis introduces a simulation-based testing method, which involves multi-sensor configurations and includes virtual environment testing with various scenarios. To deal with real-time traffic scenarios, an effective way is utilized to search for the closest scenarios. Also, the proposed method has been tested in popular autonomous vehicle platforms and simulators . To utilize the existing methodology, we perform industry platforms on Baidu Apollo and guide it to assess the quality and enhance the perception system based on a simulation-based testing platform.