Abstract
Artificial Intelligence (AI) is increasingly used in Security Operations Centers (SOCs) to support intrusion detection and threat investigation. However, accurate predictions and explainable outputs do not necessarily lead to appropriate human reliance: analysts may over-rely on incorrect recommendations or unnecessarily reject reliable ones. This work proposes a human-centric trust-calibrated framework that integrates an AI-driven Network Intrusion Detection System (NIDS), online estimation of analyst trust and cognitive load, and adaptive explanation selection within a closed-loop security interface. The NIDS combines traffic and textual threat representations to provide threat predictions, calibrated reliability estimates, and semantic evidence. Behavioral and interaction signals are processed through a probabilistic state-space model to estimate the analyst state, while a reliability-aware controller dynamically adjusts explanation depth and evidence granularity according to the Trust-Calibration Error (TCE), which captures the mismatch between inferred analyst trust and current NIDS reliability. Rather than treating trust estimation or calibration as standalone objectives, we investigate whether their integration improves operational human–AI performance under varying NIDS reliability. A controlled SOC-oriented evaluation compares the proposed trust-calibrated policy with static and trust-agnostic adaptive explanation strategies in terms of analyst decision quality, inappropriate reliance, investigation efficiency, and cognitive workload. Empirical evaluation across 2160 trials with 36 participants demonstrates that the proposed trust-calibrated policy improves decision accuracy from 81.4% to 89.3%, corresponding to an increase of 7.9 percentage points, curtails Automation Bias Rate (ABR) from 43.1% to 22.2% (31/72 to 16/72, representing a 48.4% relative reduction; GLMM-adjusted: 21.8%, a 49.4% reduction), and decreases Inappropriate Rejection Rate (IRR) from 18.8% to 11.0% (122/648 to 71/648, representing a 41.8% relative reduction; GLMM-adjusted: 18.9% to 10.9%, a 42.3% reduction), while lowering trust miscalibration error from 0.184 to 0.101. Crucially, these reliance gains are accompanied by a reduction in subjective workload from 58.4 to 48.7 on the NASA-TLX scale and a decrease in mean investigation time from 30.2 s to 28.1 s. These findings indicate that effective human-centered security interfaces should not maximize trust or explanation richness, but dynamically promote reliance that is consistent with the actual competence of the AI system.