One of the greatest challenges of representation-based face recognition is that the training samples are usually insufficient. In other words, the training set usually does not include enough samples to show varieties of high-dimensional face images caused by illuminations, facial expressions, and postures. When the test sample is significantly different from the training samples of the same subject, the recognition performance will be sharply reduced. We propose a robust kernel collaborative representation based on virtual samples for face recognition. We think that the virtual training set conveys some reasonable and possible variations of the original training samples. Hence, we design a new object function to more closely match the representation coefficients generated from the original and virtual training sets. In order to further improve the robustness, we implement the corresponding representation-based face recognition in kernel space. It is noteworthy that any kind of virtual training samples can be used in our method. We use noised face images to obtain virtual face samples. The noise can be approximately viewed as a reflection of the varieties of illuminations, facial expressions, and postures. Our work is a simple and feasible way to obtain virtual face samples to impose Gaussian noise (and other types of noise) specifically to the original training samples to obtain possible variations of the original samples. Experimental results on the FERET, Georgia Tech, and ORL face databases show that the proposed method is more robust than two state-of-the-art face recognition methods, such as CRC and Kernel CRC.