Safebench Challenge
Security and safety are critical considerations in the development and deployment of autonomous driving technology. As self-driving cars become more common, the potential risks associated with their use increase. Ensuring the security and safety of these vehicles is essential to protect passengers, other road users, and the wider public.
This challenge aims to contribute to the security and safety of the autonomous driving community. In particular, we ask the participants to design scenario-level attacks and defenses for both perception and planning modules of autonomous driving systems.
Important Dates
| Challenge Start | Mar 01, 2023 11:59PM AoE |
| Submission Open | Apr 01, 2023 11:59PM AoE |
| Submission Deadline | Jun 01, 2023 11:59PM AoE |
| Award Notification | Jun 10, 2023 11:59PM AoE |
| Top Team Presentation | Jun 19, 2023 |
Platform
Please check this website for the instruction of participating the challenge.Submission
Please wrap your submission into our docker image. Docker instructions can be found in our GitHub repository. On your Docker Hub page, please include details about how to run the docker image. Please fill in this form to submit your docker image.Tracks
We host 4 tracks in the challenge to cover the perception and planning systems of autonomous driving. The tracks are:| Track 1 (Perception Attack) | The participants are required to design adversarial textures of traffic objects (e.g., stop sign, pedestrain) to attack state-of-the-art object detection models. |
| Track 2 (Perception Defense) | The participants are required to design an robust objection model to defense adversarial attack on textures of traffic objects. |
| Track 3 (Planning Attack) | The participants are required to design traffic scenarios to reduce the safety and security of planning models (e.g., RL, rule-bsed model). |
| Track 4 (Planning Defense) | The participants are required to design safe and robust planning model to defense attacks in risky scenarios. |