To detect image spam effectively, it is necessary to analyze the image content. We do research on the local invariant features of images, and thus propose a novel method: near-duplicate image spam detecting based on CE (cross entropy), in which the SURF (Speeded up Robust Features) is used to extract the local invariant features of each image (spam and ham); then the GMM (Gaussian Mixture Models) of local invariant features are fitted. Using CE as the distance measurement between Gaussian distributions, we improve the Kmeans to cluster the GMMs since our dataset is very large. Experiments show that using CE as the distance measurement is beneficial, and the proposed method achieves better performance than some existing methods, the precision of the method can get up to 96%.