Safety evaluation and verification are challenging for the development of the Advanced Driver Assistance Systems(ADAS) or Automated Driving Functions(ADF). Real world road testing requires an unaffordable testing time. Therefore, simulations are included so that ADAS/ADF are tested and verified in a virtual environment. In the previous work, a test-case catalog has been introduced to limit the total number of test cases. Realistic parameterization of test-cases based on experimental data allows safety assessment through the estimation of unsafe conditions (e.g. collision rate) in terms of the real world traffic situation, by looking for the safety boundary, separating safe conditions from unsafe. Due to the high dimensionality of complex test cases, a rigorous grid searching is highly time consuming. Against this background, we propose an improved Input Design method based on Gaussian Process Classification algorithm to handle the safety boundary searching problem. This method allows an efficient identification of non-convex boundaries. It guarantees a good approximation of the true boundary based on limited number of simulations (or measurements).