Simulation-based virtual testing has become an essential step to ensure the safety of autonomous driving systems. Testers need to handcraft the virtual driving scenes and configure various environmental settings like surrounding traffic, weather conditions, etc. Due to the huge amount of configuration possibilities, the human efforts are subject to the inefficiency in detecting flaws in industry-class autonomous driving system. This paper proposes a coverage-driven fuzzing technique to automatically generate diverse configuration parameters to form new driving scenes. Experimental results show that our fuzzing method can significantly reduce the cost in deriving new risky scenes from the initial setup designed by testers. We expect automated fuzzing will become a common practice in virtual testing for autonomous driving systems.