Many existing source camera classification methods involve either training a classifier or computing the reference pattern noise of a camera, which means a set of images of known origins have to be pre-acquired. However, such requirement can not always be satisfied in real-world forensic applications. In this work, we propose a graph based approach that requires no extra auxiliary images nor a prior knowledge about the constitution of the image set. By formulating the classification task as a graph partitioning problem, a set of images can be classified according to their source cameras in an entirely blind way, with the number of source cameras automatically estimated. Experimental results have verified the validity of the proposed approach.