deep neural networks are vulnerable to those boundary samples. Their detection accuracy may drop significantly for these samples. In this paper, we propose an approach to generate adversarial samples based on existing layouts. These adversarial samples can help to identify the boundary of hotspot and non-hotspots. The labels of these adversarial samples can be obtained through direct litho simulation. From our observation of the litho simulation results, layouts are very sensitive to the distance between polygons. By adjusting the distances, the type of a layout may change, from a hotspot to a non-hotspot, or vice versa.