Analysis of Profile Injection Attacks against Recommendation Algorithms on Bipartite Networks

Despite their great adoption in e-commerce sites, recommender systems are still vulnerable to unscrupulous producers who try to promote their products by shilling the systems. In the past decade, network based recommendation approaches have been demonstrated to be both more efficient and of lower computational complexity than collaborative filtering methods, however as far as we know, there is rare research on the robustness of network based recommendation approaches. In this paper, we conducted a serious of experiments to examine the robustness of five typical network based recommendation algorithms. The empirical results obtained from the movielens dataset show that all the two limited knowledge shilling attacks are successful against the network based algorithms, and the bandwagon attack affects very strongly against most network based recommendation algorithms, especially the algorithms considering the preferential diffusion at the last step. One way to relieve the attack impact is to assign the algorithm a heterogeneous initial resource configuration.