Research On Detection Attribute Extraction Of Profile Injection Attack In Collaborative Filtiring Recommender System

Collaborative filtering recommendation system as one of the most widely usedpersonalized recommender systems is regarded as an effective way to solve the overloadproblem in information explosion age. However, due to the openness and user-anonymityof collaborative filtering recommendation system, there are serious hidden dangers in therecommender system. With the commercial motivations some malicious users canaltering the system’s behavior to produce some advantageous recommended results forthem by injecting some biased users’ profiles into the recommender system. So, how toensure the security of recommended system is becoming one of the hot spots ofCollaborative filtering recommendation system research. Based on the depth analysis ofcurrent researches in this field, we make a further research on the detection attribute incollaborative filtering recommender system in this paper.Aiming at the problem that the existing detection attributes are not rich enough andthe attack detection ability is also not strong enough, from a enriching detection attributeperspective, Firstly, based on a depth analysis on the characteristics of the attack profiles,aim at the fact that traditional shilling detection attributes just focus on the ratingdistribution of the attack profiles which ignore the characteristics that the attack profilesoften choose the filler items randomly, we advance the associate characteristic of theattack profile. According to the characteristic that a normal user has a high probability tosuit the associate rules between the items than the attack profile, we can detect the attackprofiles by finding out the associate rules between the items through association rulesmining algorithm and improve the ability to detect the attack users. Secondly, because ofthe difference between normal users’ profiles and the attack profiles in rating distribution,target items’ rating of the attack profile has a big different before and after attack. But thedetection attribute which is already existing now can’t reflect this characteristic of theattack profiles well. Depending on this deficiency, we advance the target-item detectionattribute which can describe the difference of the average rating of target items before andafter attack. Then we can find the target items which value of the target-item detection attribute is bigger than a threshold value. Our research in this paper can enrich thedetection attributes and improve the detection ability to the attack profiles.Finally, we give the experimental evaluations and analysis of the detection attributein our paper to confirm the effectiveness of the detection attribute, and also we make aconclusion and prospects for the further research.