Interpretable intelligent intrusion detection systems for cybersecurity in smart digital substations

• This work investigates several Explainable Artificial Intelligence (XAI) techniques in terms of their explainability, trustworthiness, complexity and their parameters in the context of Smart Digital Substations (SDSs) cybersecurity. • Existing Intrusion Detection Systems (IDSs) rely on black-box Machine/Deep Learning Models, which lack explainability and transparency, and thus the detection decisions of these IDS become untrusted and hence these systems are not secure to use. • Explainable Artificial Intelligence techniques seek to make previously untrustworthy black-box cybersecurity detection schemes in SDSs more transparent and explainable. • The findings from this work aim to guide researchers and users in the appropriate selection and implementation of XAI techniques in future AI-based IDS in SDSs, which will ensure thrust and secure use. Cybersecurity in Smart Digital Substations (SDS) is becoming more challenging in recent years due to the integration of Information and Communication Infrastructure (ICT), which makes this critical infrastructure prone to cybersecurity threats. Existing Intrusion Detection Systems (IDS) rely on black-box Machine Learning (ML) and Deep Learning (DL) based Artificial Intelligence (AI) models, which lack explainability and transparency, and thus the detection decisions of these IDS are not trusted. Explainable Artificial Intelligence (XAI) techniques seek to make these previously untrustworthy black-box detection schemes in SDS explainable. However, the implementation of XAI-based IDS in SDS is not trivial as it brings many challenges regarding which XAI technique is suitable as well as the selection of XAI parameters, impacting the interpretability, stability, and faithfulness of the explanations. To the best knowledge of the authors, this work is the first experimental review that investigates several XAI techniques in terms of their parameters, human interpretability, stability, faithfulness, and fairness in the context of SDS cybersecurity using a realistic IEC 61850 testbed and a parameter-level sensitivity analysis to provide practical analysis. This work aims to guide researchers and users in the appropriate selection and implementation of XAI techniques in future ML-based and DL-based IDS in SDS, ensuring trust and secure use.