Kernel based Support Vector Machine （SVM）does not consider inner property of training data

so classification results are usually not in optimum condition. In this paper we present a new SVM classification algorithm. The proposed method alters the kernel based on the class information of the training data

with input vectors being classified by this transformed kernel. The described algorithm can improve performance of mapping function indirectly. Simulation and experiments validate that it can improve classification performance and robustness

and reduce noise.