The Research Of Attack Detection Algorithm In Recommender System

The rapid development of e-commerce provides us with diverse choices in our life, but also makes the service information show a trend of "overload". Recommendation technology is an important tool to filter information, and it’s an effective method to solve information overload. However, the openness and sensitivity of recommender system make it easy to be attacked by the outside world. Driven by commercial interests, some vicious businessmen implant some fake user interviews deliberately into online system to influence the recommendation results. How to defense and detect the attack and ensure the security of e-commerce recommender system have become new research hotspots in the information recommendation field. This paper analyses the security research status at home and abroad of recommender system, and provides some further research for the attack detection algorithm based on collaborative filtering. The main researches are as follows:1. We analyze the basic idea and working process of collaborative filtering algorithm for understanding the strategy and relevant issues about recommendation attack. Moreover, we classify the attack model according to the grading strategy of attackers’ interviews. We classify the existing classic attack detection algorithms, and generate the corresponding attack user interviews, and implant them into the original system according to several criteria of attack model. At last, we compare the influence of different proportions of attack and filter ratio to the average prediction deviation and hitting in recommender system before and after the attack.2. We study the UnRAP unsupervised attack detection algorithm based on the Hv-score and analyze the basic ideas and implementation process of the algorithm. Based on the UnRAP detection algorithm, we cluster all users in the system beforehand and compress users’ ratings in each cluster. We improve the UnRAP algorithm by considering user groups rather than a single user, getting a group-based attack detection algorithm based on AP-UnRAP. The advantage of the improved algorithm is that it fully considers the high similarity among internal attack users and can be more accurate than base on single user’s interview when looking for the target user.3. Combined with the feature of user interviews, a mixed unsupervised attack detection algorithm AP-Mix is improved. We use user overall score behavior of user group to express each user’s grating feature by reducing the dimensionality of user’s original score matrix using PCA method and combining the dimension of the principal component information and the feature attributes of users’ interview. Then, we classify all the users into groups using an adaptive AP clustering algorithm. Finally, we calculate the average score deviation of each group(GRDMA) to find that which group the attack user is in, detecting the attack user implanted. With the combination of information, AP-Mix algorithm can represent the user’s behavior perfectly and increase the difference degree between the attack user and normal user. Based on the user group division, the detection results is better. We do not need to know any attack knowledge in advance. This algorithm realizes unsupervised detection in the true sense. At last, compared with the existing classical detection algorithm, we verify the new proposed algorithm of high efficiency.