High-knowledge shilling attack detection method based on genetic co-forest

Shilling attack is the behavior of malicious users to elevate or suppress ratings in collaborative filtering recommendation system, which can destroy the balance of recommendation and disrupt the order of competition. In previous studies, shilling attack detection techniques can be divided into three categories according to the labeling scale of samples: supervised, semi-supervised and unsupervised methods. The semi-supervised method is better at using a small number of labeled samples and a large number of unlabeled samples to improve the detection efficiency. However, in practice, it is found that the existing semi-supervised methods perform poorly in the face of high-knowledge shilling attacks such as PUA and PIA, and the generalization ability is significantly reduced when the proportion of labeled data becomes smaller. Therefore, this paper proposes a semi-supervised detection method GA-CoF. Different from the existing divergent semi-supervisors, the proposed method regards the solution set of unlabeled samples as gene sequences, the error rate as the fitness function, and the aggregation degree as the penalty term. In the experiments, we extract the user behavior characteristics of standard shilling attacks, and compare our method with four representative shilling attack detection algorithms on two datasets, MovieLens-25M and Amazon-5Cores. The results show that our method can maintain high F1 score and is not sensitive to the attack scale when facing high-knowledge shilling attacks, and performs well on real datasets. The proposed method has good module plugging ability and expansion prospects.