Page 70 - AI for Good-Innovate for Impact Final Report 2024

AI for Good-Innovate for Impact

13�2� Use case description

13�2�1 Description

Closed-Loop test the safety of autonomous system with generative and adversarial methods
in various scenarios (e.g. different environments and traffic conditions.)

On the aspect of life safety: Through closed-loop data and natural adversarial scene generation,
adversarial sample attacks, and defense reinforcement techniques, the platform can enhance
the safety and reliability of intelligent driving systems in dealing with complex environments
and threats such as attacks, effectively ensuring the safety of drivers and passengers.
On the aspect of traffic automation: The platform, through its safety simulation capabilities
and data sharing capabilities, aids in the secure implementation of intelligent driving systems,
promoting the development of intelligent transportation systems, improving road traffic
efficiency, and fostering sustainable urban development.

On the aspect of industrial development: The platform facilitates the sharing of intelligent vehicle
data and models among enterprises, breaking down data barriers, collectively enhancing
the performance and safety of intelligent driving systems. This transformation in production
relations promotes the circulation of data production factors and drives the development of
industrial systems.

13�2�2 Future Work:

Enhance capability of generating natural adversarial driving scenarios in perception & decision
tasks; Develop federated learning based sharing solutions for autonomous driving verification
scenarios across enterprises; Establish global standards on the assessment of autonomous
driving service quality; Collaborate with global experts; Global promotion and operation,
hosting or organizing related international conferences, forums, events, international youth
AI safety education, and training.

13�3� Use case requirements
• UC15-REQ01: Test the performance of autonomous vehicles in different environment
and traffic conditions. Support for a wide range of environmental simulations (e.g.
urban, highway, rural, adverse weather), and simulating various traffic scenarios (e.g.
intersections, ramps, lane-changing, pedestrian crossing), integration of multiple sensors
(e.g. radar, LIDAR, camera) for realistic perception emulation. Closed-loop simulation to
test behaviour-based vehicles from sensor inputs to control outputs.
• UC15-REQ02: Develop and integrate advanced adversarial attack algorithms to improve
scenario control policies ensuring compatibility with popular adversarial methods.
• UC15-REQ03: Generate comprehensive reports, including analytics for model
performance and evaluation, automated reporting tools to track model performance
metrics (e.g. accuracy, precision and recall), visualizing capability for analyzing model
behavior and decision-making processes, with customizable options to include specific
metrics and charts.
• UC15-REQ04: Continuously refine and enhance the performance of the autonomous
vehicle models. Support for iterative developing and rapid prototyping. Analyzing
simulation results and identifying area for improvement. Integration with version contro
systems to manage model iterations and updates. Support for parallel testing of multiple
model versions to compare performance.

54