A trolling attack detection algorithm based on popularity classification features

Abstract

The invention relates to a shilling attack detection algorithm based on popularity classification features. The algorithm first collects the statistics of the scores of items given by users, and creates a user score matrix; the statistics of the item popularity of the items are then collected; user popularity vectors are then determined; classification feature values (MUD, RUD and QUD) based on popularity are then calculated; a classifier is then created, finally, elements in a user popularity vector of a new user are inputted into the classifier, and thereby whether the new user is a normal user or a false user can be determined. The detection algorithm provided by the invention has a good judgment effect on user classes, and has excellent shilling attack detection performance on both simple random attacks, evaluation attacks and bandwagon attacks and attacks in confusion jamming, moreover, calculation cost is low, and detection time is shorter.

Description

technical field [0001] The invention relates to the field of information security, in particular to a trolling attack detection algorithm based on popularity classification features. Background technique [0002] Recommender systems are an important tool in the field of e-commerce to select potentially interesting items for users. Collaborative filtering is a technology widely used in recommendation systems. This method finds the most similar users as the nearest neighbors for the target users, and uses the purchase information of the nearest neighbors to generate recommendation results. This working mode is very effective in practice, but it is vulnerable to Shilling attacks. Attackers inject certain false profiles into the nearest neighbors of normal users to interfere with the recommendation results of the recommendation system, thereby increasing or reducing the recommendation frequency of the target item, which are called Push and Nuke Attacks, respectively. How to pre...

AI Technical Summary

Problems solved by technology

Most of the classification features currently used are related to users’ ratings on items, that is, to find corresponding detection indicators from the different ways of normal users and fake users’ ratings on items. There are two problems with this method of detection: (1) The scoring methods of some normal users and fake users are similar, which may easily lead to misjudgment of such normal users; (2) Most of the actual attacks are confused, such as not rating the highest (low) score for the target item but rating High (low) score or adding a random number as noise interference on the basis of the original score, so that the current detection indicators are difficult to cope with various changes in the attack method

Images