工業控制系統廣泛應用於製造、能源、水處理與交通等關鍵基礎設施場域，其安全事件不僅可能造成資訊外洩，更可能導致系統狀態中斷、設備損壞與人員傷亡風險。近年來，多模態入侵偵測系統因能同時結合不同資訊來源而逐漸受到重視；然而，多模態深度學習模型通常具有較高之參數量與運算成本，使其難以部署於資源受限之邊緣設備。
為此，本研究建立一套面向工業控制系統之多模態入侵偵測基準模型，結合實體模態（physical modality）之實體系統狀態與致動器訊號，以及由同一底層資料衍生建構之行為模態（behavioral modality）特徵，並導入SHAP（SHapley Additive exPlanations）分析模態與特徵貢獻。基於可解釋性分析結果，本研究提出階層式 SHAP 引導結構化剪枝方法（Hierarchical SHAP-Guided Structured Pruning, HGSP），先依模態貢獻分配不同分支之剪枝強度，再結合特徵重要度與通道層啟動值重要度進行濾波器保留決策，並於剪枝後進行微調，以在壓縮模型之同時保留重要辨識能力。
本研究以 SWaT 整合後時間序列資料表進行實驗，並同時採用剪枝比例比較與固定部署預算比較兩種設計。實驗結果顯示，SHAP 分析證明behavioral modality 對最終分類具有實質輔助價值，而非僅為冗餘資訊；在不同剪枝比例下，HGSP 於中高剪枝區間仍可維持相對穩定之 F1-score 與 Attack Recall，且推論延遲隨剪枝比例增加而下降。進一步在固定約260 KB之部署預算下比較時，HGSP 取得最高之 F1-score（0.8339）與最低之推論延遲（599.73ms），整體表現優於非結構化幅值剪枝、結構化幅值剪枝與結構化啟動值剪枝三種對照方法。研究結果顯示，以模態貢獻差異引導差異化結構化剪枝，可作為工業控制系統 IDS 輕量化設計之可行方向，並具備進一步朝向邊緣部署發展之潛力。

Industrial Control Systems (ICS) face increasing cyber threats, yet multimodal intrusion detection models are often too large for resource-constrained edge devices. This study develops a dual-branch multimodal anomaly detection baseline model integrating a physical modality and a behavioral modality derived from the SWaT (Secure Water Treatment) dataset, and uses SHAP (SHapley Additive exPlanations) to quantify each modality's contribution. We propose Hierarchical SHAP-Guided Structured Pruning (HGSP), which allocates pruning capacity according to modality- and channel-level contribution rather than uniform compression. Results show that the behavioral modality contributes approximately 32% of overall SHAP importance, confirming its complementary value, and that HGSP achieves the highest F1-score and lowest inference latency among compared methods under a fixed deployment budget of approximately 260 KB.

Adamczewski, K., Li, Y., and Van Gool, L. (2024). Shapley pruning for neural network compression. arXiv preprint arXiv:2407.15875.
Ancona, M., Öztireli, C., and Gross, M. (2020). Shapley value as principled metric for structured network pruning. arXiv preprint arXiv:2006.01795.
Anthi, E., Williams, L., Burnap, P., and Jones, K. (2021). A three-tiered intrusion detection system for industrial control systems. Journal of Cybersecurity, 7(1).
Benaddi, H., Jouhari, M., Laamech, N., Motii, A., and Ibrahimi, K. (2025). Lightweight intrusion detection in IoT via SHAP-guided feature pruning and knowledge-distilled Kronecker networks. arXiv preprint arXiv:2512.19488.
Bhushan, C. M., Koppuravuri, P., Prasanthi, N., Gazi, F., Hussain, M. M., Abdussami, M., Devi, A. A., and Faizi, J. (2025). Deploying TinyML for energy-efficient object detection and communication in low-power edge AI systems. Scientific Reports, 15:44299.
Cheng, H., Zhang, M., and Shi, J. Q. (2023). A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommendations. arXiv preprint arXiv:2308.06767.
Cherepanov, A., Breza, J., and Calvet, J. (2022). Industroyer2: Industroyer reloaded. Technical report, ESET.
Fan, Y., Xu, W., Wang, H., Wang, J., and Guo, S. (2023). PMR: Prototypical modal rebalance for multimodal learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20029–20038.
Goh, J., Adepu, S., Junejo, K. N., and Mathur, A. (2016). A dataset to support research in the design of secure water treatment systems. In Critical Information Infrastructures Security, pages 88–99. Springer.
Jahromi, A. N., Karimipour, H., Halabi, T., Zhu, Y., and Gadekallu, T. R. (2025). Multimodal game-theoretic cyber-attack projection in industrial control systems. IEEE Transactions on Consumer Electronics, 71(2):5801–5811.
Kiflay, M., Tsokanos, G., Fazlali, H., and Kirner, R. (2024). Network intrusion detection leveraging multimodal features. Array, 24:100349.
Lee, R. M., Assante, M. J., and Conway, T. (2016). Analysis of the cyber attack on the Ukrainian power grid: Defense use case. Technical report, Electricity Information Sharing and Analysis Center (E-ISAC) and SANS Industrial Control Systems.
Li, C., Li, H., Xue, M., Fang, G., Zhou, S., Feng, Z., Wang, H., Song, M., and Song, J. (2024). A comprehensive study of structural pruning for vision models. arXiv preprint arXiv:2406.12315.
Li, H., Kadav, A., Durdanovic, I., Samet, H., and Graf, H. P. (2017). Pruning filters for efficient ConvNets. In International Conference on Learning Representations (ICLR).
Lu, Y., Jiang, B., Liu, N., Li, Y., Chen, J., Zhang, Y., and Wan, Z. (2024). CrossPrune: Cooperative pruning for camera–LiDAR fused perception models of autonomous driving. Knowledge-Based Systems, 289:111522.
Lundberg, S. M. and Lee, S.-I. (2017). A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, volume 30.
Makrakis, G. M., Kolias, C., Kambourakis, G., Rieger, C., and Benjamin, J. (2021). Industrial and critical infrastructure security: Technical analysis of real-life security incidents. IEEE Access, 9:155222–155252.
Manu, G. J. and Kunte, R. S. R. (2024). FDI attack detection in industrial control systems using 1d-CNN: A comparative study on SWaT, WADI, and MATPOWER datasets. Journal of Information Systems Engineering and Management, 9(4s).
Maynard, P., McLaughlin, K., and Sezer, S. (2020). Decomposition and sequential AND analysis of known cyber-attacks on critical infrastructure control systems. Journal of Cybersecurity, 6(1).
Merkle, F., Weber, D., Schöttle, P., Schlögl, S., and Nocker, M. (2025). Less is more: The influence of pruning on the explainability of CNNs. IEEE Access.
O'Meara, K., Graham, M., and Ahlers, C. (2024). Impact of FrostyGoop ICS malware on connected OT systems. Technical report, Dragos, Inc.
Oyedotun, S. A., Oise, G. P., and Ozobialu, C. E. (2025). Towards intelligent cybersecurity in SCADA and DCS environments: Anomaly detection using multimodal deep learning and explainable AI. Journal of Science Research and Reviews, 2(3):20–31.
Qazi, E. U. H., Almorjan, A., and Zia, T. (2022). A one-dimensional convolutional neural network (1d-CNN) based deep learning system for network intrusion detection. Applied Sciences, 12(16):7986.
Sallam, S., El Barachi, M., and Li, N. (2026). Intrusion detection on the internet of things: A comprehensive review and gap analysis toward real-time, lightweight, adaptive, and autonomous security. IoT, 7(1):16.
Schizas, N., Karras, A., Karras, C., and Sioutas, S. (2022). TinyML for ultra-low power AI and large scale IoT deployments: A systematic review. Future Internet, 14(12):363.
Shukla, A. and Tapas, N. (2026). Explainability-inspired layer-wise pruning of deep neural networks for efficient object detection. arXiv preprint arXiv:2602.14040.
Singh, R. and Gill, S. S. (2023). Edge AI: A survey. Internet of Things and Cyber-Physical Systems, 3:71–92.
Wu, T., Song, C., and Zeng, P. (2023). Model pruning based on filter similarity for edge device deployment. Frontiers in Neurorobotics, 17:1132679.
Xu, C., Zhan, Y., Wang, Z., and Yang, J. (2025). Multimodal fusion based few-shot network intrusion detection system. Scientific Reports, 15:21986.
Yuan, T., Yang, S., Liu, B., and Tang, Y. (2025). Explainable CNN filter pruning based on Shapley value approximation in view of game theory. Expert Systems with Applications, 270:130232.