Scholarlink

This study presents TGA-Greedy (TGA-Gre), a novel adversarial attack strategy that exploits vulnerabilities in Deep Dynamic Network Embedding (DDNE) models. Unlike traditional traversal-based methods such as TGA-Traversal (TGA-Tra), TGA-Gre adopts a gradient-based approach to strategically perturb dynamic network snapshots by targeting critical nodes with high centrality measures. TGA-Gre effectively leverages gradient information to achieve significant disruption with minimal modifications, enhancing computational efficiency and overall impact. Experimental evaluations demonstrate the effectiveness of the TGA-Gre attack, achieving an Attack Success Rate (ASR) of 70.17% with an Average Modification Links (AML) of 10.00, while reducing DDNE prediction accuracy from 50.80% to 47.77%. This research underscores the vulnerability of DDNE models to targeted attacks and highlights the importance of adversarial robustness in dynamic network applications, establishing a foundation for developing future defenses against such threats.

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

Evaluating the Robustness of Graphs Against Link Prediction Attack Threats

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tran Nguyen,
  • Quang-Vinh Dang

摘要

This study presents TGA-Greedy (TGA-Gre), a novel adversarial attack strategy that exploits vulnerabilities in Deep Dynamic Network Embedding (DDNE) models. Unlike traditional traversal-based methods such as TGA-Traversal (TGA-Tra), TGA-Gre adopts a gradient-based approach to strategically perturb dynamic network snapshots by targeting critical nodes with high centrality measures. TGA-Gre effectively leverages gradient information to achieve significant disruption with minimal modifications, enhancing computational efficiency and overall impact. Experimental evaluations demonstrate the effectiveness of the TGA-Gre attack, achieving an Attack Success Rate (ASR) of 70.17% with an Average Modification Links (AML) of 10.00, while reducing DDNE prediction accuracy from 50.80% to 47.77%. This research underscores the vulnerability of DDNE models to targeted attacks and highlights the importance of adversarial robustness in dynamic network applications, establishing a foundation for developing future defenses against such threats.