A method for constructing a poisoning sample based on user classification

By classifying users in the recommendation system and constructing a proxy model with dynamic weights, fake users are generated, which solves the problem of insufficient user feature differentiation in existing technologies, achieves stronger data poisoning attacks and lower attack costs, and provides a defense strategy.

CN115935188BActive Publication Date: 2026-07-07NINGBO UNIV
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
NINGBO UNIV
Filing Date
2022-12-16
Publication Date
2026-07-07

AI Technical Summary

Technical Problem

Existing data poisoning attack techniques based on deep learning recommendation systems fail to effectively distinguish user characteristics, resulting in unsatisfactory promotion effects, high attack costs, and a lack of targeted defense measures.

Method used

By classifying users of the recommendation system and defining vulnerable and robust users, a proxy model is constructed using dynamic weights to simulate the maximum poisoning state of vulnerable users. Fake users are then generated and added to the dataset, reducing attack costs and improving attack effectiveness.

Benefits of technology

It enhances the offensiveness of data poisoning attacks, reduces attack costs, provides new ideas for the defense of recommendation systems, improves the poisoning effect on vulnerable users, and reduces the poisoning impact on robust users.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115935188B_ABST
    Figure CN115935188B_ABST
Patent Text Reader

Abstract

The application discloses a kind of based on user classification's poisoning sample construction method, comprising the following steps: define vulnerable user and robust user in system;Define loss function, i.e. the dynamic weight of each user, construct agent model, simulate system vulnerable user as far as possible the state of poisoning to obtain false user candidate interaction article candidate set, and reduce recommendation bias by initializing false user to improve the credibility of candidate interaction article;Define the selection probability of each article, obtain the final score of all articles, and select the interaction article of false user according to score from high to low.The application enhances the attack of data poisoning attack based on deep learning recommendation system, reduces the attack cost, and provides a thought for the defense of recommendation system to data poisoning attack.
Need to check novelty before this filing date? Find Prior Art