Explainable AI in Data Poisoning Threat Models Across the CIA Triad: A Smart Grid Case Study

eXplainable Artificial Intelligence (XAI) techniques make models more interpretable, but this can help malicious actors to better achieve their goals. In this paper, we systematically apply data poisoning to the most important features according to distinct XAI methods, and evaluate its impact on learning-based model architectures trained on realistic Smart Grid (SG) datasets. Our approach involves poisoning top-ranked features to determine which XAI method-or combination thereof-provides the most actionable insights to improve data poisoning attacks in different settings. Furthermore, we perform a practical problemspace attack via Global Navigation Satellite System (GNSS) time spoofing, where time-related features identified as most critical by XAI techniques are subsequently exploited, leading to a decrease in Intrusion Detection System (IDS) performance. This work contributes to the community by offering new insights into the vulnerability of critical infrastructure systems to sophisticated -yet realistic- adversarial attacks.