The application of kernel functions can solve the problem of non-linear image data so that the data can be linearly separable with a hyperplane by mapping the input space to the feature space to increase its dimensions. This article will discuss the improvement in recognition accuracy obtained by implementing multiple kernels in a PCA based program using linear, polynomial, and gaussian kernels for facial recognition with illumination variations. The matching or recognition process is carried out using the SVM method. Improvements obtained from the application of multiple kernels will be compared with the implementation of a single kernel and see how much improvement of the accuracy. Based on the results of the implementation of multiple kernels, the average improvement in accuracy obtained from the face recognition results with illumination variations is 10.5% compared to a single kernel.