Scholarlink

<p>Trust evaluation is an essential element in the decision-making process for implementing cybersecurity measures. Trust evaluation has been extensively utilized in various domains to achieve disparate objectives. In the construction of next-generation communication networks, access risk can be effectively mitigated by evaluating the trustworthiness of network nodes. Evidence for trust evaluation should be multi-source and comprehensive. However, current research on trust evaluation does not account for the interdependence among multidimensional trust evidence and fails to consider the stealthy nature of trust-related attacks. This makes it difficult to ensure a dynamic and secure evaluation process. To address these issues, we propose a multidimensional trust fusion evaluation model based on Bi-LSTM and spatio-temporal cross-attention, which learns the interactions of trust evidence with robust perception capabilities. A cross-attention mechanism is utilized to dynamically weight the contributions of trust evidence extracted from different sources. A robust filtering layer is incorporated into the embedded representation module, thereby ensuring resilience against various stealthy behavioral trust attacks. Furthermore, in the context of temporal contextual trust information, local features are generated by Bi-LSTM to address both short-term fluctuations and long-range dependencies. Subsequently, a gating mechanism is introduced to facilitate the fusion of global and local features. The system can adaptively balance local and global features. Extensive experiments on real-world datasets demonstrate that compared to state-of-the-art trust prediction methods, the proposed model achieves superior performance, with the F1-micro score improved by approximately 7% in standard evaluations. Furthermore, the model exhibits exceptional robustness; under severe collaborative attack scenarios, it outperforms the strongest baseline methods by a relative margin of over 25%, maintaining high reliability where other models degrade significantly.</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

ST-CATrust: multidimensional trust fusion evaluation via spatio-temporal graph neural network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月30日

查看全文
  • Zihao Wang,
  • Yuxiang Hu,
  • Xu Feng,
  • Pengshuai Cui,
  • Le Tian,
  • Jinchuan Pei

摘要

Trust evaluation is an essential element in the decision-making process for implementing cybersecurity measures. Trust evaluation has been extensively utilized in various domains to achieve disparate objectives. In the construction of next-generation communication networks, access risk can be effectively mitigated by evaluating the trustworthiness of network nodes. Evidence for trust evaluation should be multi-source and comprehensive. However, current research on trust evaluation does not account for the interdependence among multidimensional trust evidence and fails to consider the stealthy nature of trust-related attacks. This makes it difficult to ensure a dynamic and secure evaluation process. To address these issues, we propose a multidimensional trust fusion evaluation model based on Bi-LSTM and spatio-temporal cross-attention, which learns the interactions of trust evidence with robust perception capabilities. A cross-attention mechanism is utilized to dynamically weight the contributions of trust evidence extracted from different sources. A robust filtering layer is incorporated into the embedded representation module, thereby ensuring resilience against various stealthy behavioral trust attacks. Furthermore, in the context of temporal contextual trust information, local features are generated by Bi-LSTM to address both short-term fluctuations and long-range dependencies. Subsequently, a gating mechanism is introduced to facilitate the fusion of global and local features. The system can adaptively balance local and global features. Extensive experiments on real-world datasets demonstrate that compared to state-of-the-art trust prediction methods, the proposed model achieves superior performance, with the F1-micro score improved by approximately 7% in standard evaluations. Furthermore, the model exhibits exceptional robustness; under severe collaborative attack scenarios, it outperforms the strongest baseline methods by a relative margin of over 25%, maintaining high reliability where other models degrade significantly.