Abstract
<title>Abstract</title> <p>Explainable artificial intelligence (XAI) helps operators interpret network-intrusion alerts, but generating explanations consumes compute, introduces queueing delay, accumulates explanation debt, and risks exposing defensive classifier decision boundaries. Current XAI-IDS research concentrates almost exclusively on semantic interpretability, ignoring these operational service-management challenges under resource constraints and adversarial explanation-demand inflation. This paper introduces XAI-Gate, a novel, detector-agnostic online service-governance framework that dynamically selects among seven explanation actions - suppress, coarse, delay, full local, offload, audit-only, or redact - based on packet pressure, explanation debt, offload belief, suspicion state, and leakage-budget constraints. We establish XAI-SurfaceBench, a rigorous stress-test benchmark anchored on KDDCup99, UNSW-NB15, and TON_IoT score-stream datasets to evaluate performance across nine traffic/uncertainty regimes. In the default adversarial explanation-demand inflation regime, XAI-Gate maintains absolute exposure feasibility (zero budget violations) while achieving a high-risk explanation coverage of 0.800+/-0.000 and reducing mean explanation debt by 57.3% compared to conventional threshold policies. Our findings demonstrate that XAI-Gate serves as a robust control plane that balances the operational frontier between packet QoS, operator interpretability, and information exposure without requiring classifier re-training.</p>