| 電子情報通信学会 研究会発表申込システム 講演論文 詳細 |
技報閲覧サービス [ログイン] 技報アーカイブ |
| トップに戻る | 前のページに戻る | [Japanese] / [English] |
| 講演抄録／キーワード | ||
| 講演名 | 2026-03-06 10:45 Efficient Training and Explainable GNNs for IoT Intrusion Detection via Modified Neighbor Sampling ○Rafif Reyhandhia Yusa・Ryo Yamamoto・Satoshi Ohzahata（UEC）  NS2025-305 |
|
| 抄録 | （和） | While the widespread adoption of Internet of Things (IoT) has provided transformative benefits across various sectors, its inherent vulnerabilities continue to pose significant security risks. With threats continuously evolving, traditional rule-based approaches are becoming more difficult to manage and scale. Simultaneously, conventional Machine Learning (ML) approaches struggle to capture the inherent complex relationship between devices in IoT network communication, suffer from performance degradation due to class imbalance in the training data, and remain opaque in their decision making–commonly referred as the "black-box" problem. To address these challenges, this research proposes an interpretable Network Intrusion Detection System (NIDS) that captures the spatial and topological information of the network by mapping it into a graph structure. By utilizing GraphSAGE for inductive learning, interpretable explanations are generated post-hoc using the Captum library to provide insights into the detection result. Furthermore, the Layer-Neighbor Sampling algorithm and Class-Balanced Loss calculation are introduced to optimize training efficiency and help mitigate the effects of imbalanced data. Experimental results across two benchmark NIDS datasets demonstrate that the proposed framework improves detection rates for minority attack classes and balances Macro F1-scores, significantly reduces training time, and provides insights into the features that drive the model predictions. |
| （英） | While the widespread adoption of Internet of Things (IoT) has provided transformative benefits across various sectors, its inherent vulnerabilities continue to pose significant security risks. With threats continuously evolving, traditional rule-based approaches are becoming more difficult to manage and scale. Simultaneously, conventional Machine Learning (ML) approaches struggle to capture the inherent complex relationship between devices in IoT network communication, suffer from performance degradation due to class imbalance in the training data, and remain opaque in their decision making–commonly referred as the "black-box" problem. To address these challenges, this research proposes an interpretable Network Intrusion Detection System (NIDS) that captures the spatial and topological information of the network by mapping it into a graph structure. By utilizing GraphSAGE for inductive learning, interpretable explanations are generated post-hoc using the Captum library to provide insights into the detection result. Furthermore, the Layer-Neighbor Sampling algorithm and Class-Balanced Loss calculation are introduced to optimize training efficiency and help mitigate the effects of imbalanced data. Experimental results across two benchmark NIDS datasets demonstrate that the proposed framework improves detection rates for minority attack classes and balances Macro F1-scores, significantly reduces training time, and provides insights into the features that drive the model predictions. |
|
| キーワード | （和） | Graph Neural Network / Intrusion Detection / IoT Security / XAI / / / / |
| （英） | Graph Neural Network / Intrusion Detection / IoT Security / XAI / / / / | |
| 文献情報 | 信学技報, vol. 125, no. 385, NS2025-305, pp. 497-502, 2026年3月. | |
| 資料番号 | NS2025-305 | |
| 発行日 | 2026-02-25 (NS) | |
| ISSN | Online edition: ISSN 2432-6380 | |
| 著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します．(許諾番号：10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) | |
| PDFダウンロード | NS2025-305 |
|
| 研究会情報 | |
| 研究会 | IN NS |
| 開催期間 | 2026-03-04 - 2026-03-06 |
| 開催地（和） | 沖縄県市町村自治会館 |
| 開催地（英） | Okinawa-Ken Shichoson Jichi Kaikan |
| テーマ（和） | 一般 |
| テーマ（英） | General |
| 講演論文情報の詳細 | |
| 申込み研究会 | NS |
| 会議コード | 2026-03-IN-NS |
| 本文の言語 | 英語 |
| タイトル（和） | |
| サブタイトル（和） | |
| タイトル（英） | Efficient Training and Explainable GNNs for IoT Intrusion Detection via Modified Neighbor Sampling |
| サブタイトル（英） | |
| キーワード(1)（和/英） | Graph Neural Network / Graph Neural Network |
| キーワード(2)（和/英） | Intrusion Detection / Intrusion Detection |
| キーワード(3)（和/英） | IoT Security / IoT Security |
| キーワード(4)（和/英） | XAI / XAI |
| キーワード(5)（和/英） | / |
| キーワード(6)（和/英） | / |
| キーワード(7)（和/英） | / |
| キーワード(8)（和/英） | / |
| 第1著者 氏名（和/英/ヨミ） | Rafif Reyhandhia Yusa / Rafif Reyhandhia Yusa / |
| 第1著者 所属（和/英） | 電気通信大学 (略称： 電通大) The University of Electro-Communications (略称： UEC) |
| 第2著者 氏名（和/英/ヨミ） | 山本 嶺 / Ryo Yamamoto / |
| 第2著者 所属（和/英） | 電気通信大学 (略称： 電通大) The University of Electro-Communications (略称： UEC) |
| 第3著者 氏名（和/英/ヨミ） | 大坐畠 智 / Satoshi Ohzahata / |
| 第3著者 所属（和/英） | 電気通信大学 (略称： 電通大) The University of Electro-Communications (略称： UEC) |
| 講演者 | 第1著者 |
| 発表日時 | 2026-03-06 10:45:00 |
| 発表時間 | 25分 |
| 申込先研究会 | NS |
| 資料番号 | NS2025-305 |
| 巻番号（vol） | vol.125 |
| 号番号（no） | no.385 |
| ページ範囲 | pp.497-502 |
| ページ数 | 6 |
| 発行日 | 2026-02-25 (NS) |
[研究会発表申込システムのトップページに戻る]