Online auctions have become very popular over the last few years. This popularity is evidenced by the explosive growth of online auction sites with millions of users buying and selling goods from all over the world. However, this rapid growth of online auctions has also led to a corresponding increase in online frauds. While collusive auction frauds are not as common as other types of online frauds, they are more dangerous because they are more difficult to detect and often result in larger financial losses. In recent years, a number of techniques have been proposed to detect collusive frauds in online auction networks. While all the techniques have shown promising results, they often suffer from slow convergence or low detection performance. In this paper, we address these shortcomings by presenting CAFD, a novel anomaly detection technique that combines one-class classification and collective classification to detect collusive auction frauds. CAFD uses a one-class classifier to calculate an anomaly score for each unla-beled user. It also models the auction interactions between different users as a pairwise Markov random field (MRF) and applies belief propagation to the MRF to revise those anomaly scores. The results of our experiments show that CAFD is able to detect different types of collusive auction frauds with a low false positive rate and a reasonable detection time.