Scholarlink

<p>With the proliferation of online platforms, user reviews have become pivotal in shaping decision-making processes, rendering the timely and accurate detection of fraudulent reviews a critical imperative. However, existing detection methodologies predominantly employ offline paradigms, which are limited to analyzing historical reviews and often fail to identify sophisticatedly crafted deceptive content. To address these dual challenges, this paper proposes DAGCN, a Dynamic Adversarial Graph Collaborative Network. The model first constructs localized subgraphs by modeling existing offline data. Upon the emergence of new reviews, the Dynamic Subgraph Cache module dynamically integrates incoming nodes into appropriate subgraphs, enabling localized updates and forward propagation to minimize detection latency and achieve real-time processing. Concurrently, the Adversarial Robustness Enhancer module fortifies the model against malicious data manipulation and adversarial attacks by injecting perturbations into node features and adjacency matrices to generate adversarial training samples. Extensive experimental evaluations demonstrate that DAGCN significantly outperforms state-of-the-art benchmarks in detection accuracy while maintaining an average inference latency that satisfies real-time requirements for online systems. Our code is available at: <a href="https://github.com/w1019251220/DAGCN">https://github.com/w1019251220/DAGCN</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dynamic Adversarial GNN for Real-Time Fraud Detection on Online Platforms

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月7日

查看全文
  • Zhanming Wan,
  • Kejin Li,
  • Yingpei Wu

摘要

With the proliferation of online platforms, user reviews have become pivotal in shaping decision-making processes, rendering the timely and accurate detection of fraudulent reviews a critical imperative. However, existing detection methodologies predominantly employ offline paradigms, which are limited to analyzing historical reviews and often fail to identify sophisticatedly crafted deceptive content. To address these dual challenges, this paper proposes DAGCN, a Dynamic Adversarial Graph Collaborative Network. The model first constructs localized subgraphs by modeling existing offline data. Upon the emergence of new reviews, the Dynamic Subgraph Cache module dynamically integrates incoming nodes into appropriate subgraphs, enabling localized updates and forward propagation to minimize detection latency and achieve real-time processing. Concurrently, the Adversarial Robustness Enhancer module fortifies the model against malicious data manipulation and adversarial attacks by injecting perturbations into node features and adjacency matrices to generate adversarial training samples. Extensive experimental evaluations demonstrate that DAGCN significantly outperforms state-of-the-art benchmarks in detection accuracy while maintaining an average inference latency that satisfies real-time requirements for online systems. Our code is available at: https://github.com/w1019251220/DAGCN.