Scholarlink

<p>The fast adoption of cyber-physical devices in new smart grids has significantly increased the attack surface rendering the conventional intrusion detection systems (IDS) ineffective to counter cunning, zero-day, and coordinated attacks. Current GNN-based detectors make use of pairwise graph edges, and do not take into account higher-order multi-node dependencies inherent to power substations and multi-terminal topologies. In the current paper, HSTGAT-IDS, a Hypergraph Spatio-Temporal Graph Attention Network in detecting zero-day attacks in smart grids, is proposed. HSTGAT-IDS models the cyber-physical grid as a time-varying hypergraph where hyperedges encode group relationships (subgrid clusters, relay-breaker assemblies and time k-NN functional dependencies). Multi-head graph attention is used on top of hyperedge neighbourhood and on top of time sequences and an encoder-decoder reconstruction mechanism trained on normal data only can supply zero-day anomaly scoring. The results on the MSU-ORNL Power System Attack Dataset and the IEEE 14/118/300-bus benchmark systems show a binary detection accuracy of 99.31%, multi-class detection accuracy of 99.08%, and zero-day detection rate of 91.4%, and leads to an average F1-score of 4.7% higher than the best baseline (GraphKAN). The additive, complementary effect of each of the proposed components is proven by ablation studies.</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

A Hypergraph Spatio-Temporal Graph Attention Network for Zero-Day Attack Detection in Smart Grids

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • M. Sudha,
  • R. Reenadevi,
  • R. Jennie Bharathi,
  • B. Ardly Melba Reena,
  • Akila Venkatraman,
  • Jayavarapu Karthik

摘要

The fast adoption of cyber-physical devices in new smart grids has significantly increased the attack surface rendering the conventional intrusion detection systems (IDS) ineffective to counter cunning, zero-day, and coordinated attacks. Current GNN-based detectors make use of pairwise graph edges, and do not take into account higher-order multi-node dependencies inherent to power substations and multi-terminal topologies. In the current paper, HSTGAT-IDS, a Hypergraph Spatio-Temporal Graph Attention Network in detecting zero-day attacks in smart grids, is proposed. HSTGAT-IDS models the cyber-physical grid as a time-varying hypergraph where hyperedges encode group relationships (subgrid clusters, relay-breaker assemblies and time k-NN functional dependencies). Multi-head graph attention is used on top of hyperedge neighbourhood and on top of time sequences and an encoder-decoder reconstruction mechanism trained on normal data only can supply zero-day anomaly scoring. The results on the MSU-ORNL Power System Attack Dataset and the IEEE 14/118/300-bus benchmark systems show a binary detection accuracy of 99.31%, multi-class detection accuracy of 99.08%, and zero-day detection rate of 91.4%, and leads to an average F1-score of 4.7% higher than the best baseline (GraphKAN). The additive, complementary effect of each of the proposed components is proven by ablation studies.