主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2025, Vol. 33 ›› Issue (2): 95-104.doi: 10.16381/j.cnki.issn1003-207x.2022.0170cstr: 32146.14.j.cnki.issn1003-207x.2022.0170

• • 上一篇    下一篇

基于重要性最大化与社区划分的图神经网络推荐系统对抗攻击方法

柴一栋1,2, 刘昊鑫1,2, 姜元春1,2,4(), 刘业政1,3,4   

  1. 1.合肥工业大学管理学院,安徽 合肥 230009
    2.网络空间行为与管理安徽省哲学社会科学;重点实验室,安徽 合肥 230009
    3.过程优化与智能决策教育部重点实验室,安徽 合肥 230009
    4.大数据流通与交易技术国家工程实验室,上海 201203
  • 收稿日期:2022-01-24 修回日期:2022-07-03 出版日期:2025-02-25 发布日期:2025-03-06
  • 通讯作者: 姜元春 E-mail:ycjiang@hfut.edu.cn
  • 基金资助:
    国家自然科学专项基金项目(72342011);国家自然科学基金优青项目(72322019);国家自然科学基金面上项目(72171071);国家自然科学基金重点项(91846201);国家自然科学基金青年项目(72101079)

Graph Neural Network Recommendation System Adversarial Attack Method Based on Importance Maximization and Community Partition

Yidong Chai1,2, Haoxin Liu1,2, Yuanchun Jiang1,2,4(), Yezheng Liu1,3,4   

  1. 1.School of management,Hefei University of technology,Hefei 230009,China
    2.Key Laboratory of Philosophy and Social Sciences for Cyberspace Behaviour and Management,Hefei 230009,China
    3.Key Laboratory of Process Optimization and Intelligent Decision-making,Ministry of Education,Hefei 230009,China
    4.National Engineering Laboratory for Big Data Distribution and Exchange Technologies,Shanghai 201203,China
  • Received:2022-01-24 Revised:2022-07-03 Online:2025-02-25 Published:2025-03-06
  • Contact: Yuanchun Jiang E-mail:ycjiang@hfut.edu.cn

摘要:

推荐系统的安全性是目前社会关注的热点问题。由于对抗攻击能够评估并提升推荐系统对抗鲁棒性,推荐系统的对抗攻击成为重要的研究问题。针对图卷积神经网络在推荐系统的广泛使用,本文提出一种面向图神经网络推荐系统的对抗攻击方法。该方法基于图结构干扰的思想,构建对抗攻击的优化模型。然而,该优化模型难以直接求解,因此提出了基于重要性最大化的间接求解方法。在此基础上,针对大规模用户-物品异质图,本文进一步提出社区划分的攻击策略,通过考虑社区规模进一步提升节点重要性评估的效果与效率。基于真实数据的实验表明,本文所提方法具有更好的对抗攻击效果,能够在不改变图结构的同时,有效规避现有推荐系统防御策略。所提方法为设计对抗鲁棒性更强的推荐系统提供重要依据。

关键词: 推荐系统, 图神经网络, 对抗攻击, 重要性最大化, 社区划分

Abstract:

The security of recommendation system is a hot topic in society. Since adversarial attacks can evaluate and improve the adversarial robustness of recommendation systems, adversarial attacks of recommendation systems have become an important research issue. In view of the widespread use of graph convolutional neural network in recommendation system, we propose an adversarial attack method for graph neural network recommendation system. Based on the idea of structure perturbation, an optimization model is constructed for adversarial attacks. However, the optimization model is difficult to solve directly, so an indirect method based on importance maximization is proposed. On this basis, for the large-scale user-item heterogeneous graph, the attack strategy of community division is proposed, and the effect and efficiency of node importance assessment are improved by considering the community size. Experiments based on the real public dataset show that the proposed method has a better adversarial attack result, and can effectively avoid the existing defense strategy of recommendation system without changing the heterogeneous graph structure of user items. The proposed method provides an important basis for designing a more robust recommendation system.

Key words: recommendation system, graph neural network, adversarial attack, importance maximization, community partition

中图分类号: