An unsupervised detection method for shilling attacks based on deep learning and community detection

In the detection methods for shilling attacks, the supervised methods require labeled samples to train the classifiers. Due to lack of the labeled sample profiles in real scenarios, the applicability of supervised detection method is restricted. For unsupervised methods, the prior knowledge is often required to guarantee the detection accuracy. To break the traditional limitations, we present an unsupervised method to detect various shilling profiles from reconstructed user–user graph based on deep learning and community detection. Firstly, we construct the user–user graph, whose edge weight is calculated by the similarity of user’s behaviors. Secondly, the stacked denoising autoencoders are used to extract the robust graph features at different scales with different corruption rates. Based on the features at different scales, we generate multiple clustering results and reconstruct the user–user graph by evidence accumulation method. Thirdly, the community detection is carried out by using the persistence optimization algorithm. Extensive experiments on two datasets illustrate that our proposed method has better performance than some baseline detectors for detecting the simulated attacks and actual attacks.