CALIBURN: A Regime-Sensitivity Study of Operationally Calibrated Streaming Intrusion Detection

Streaming network intrusion detection systems must process flows continuously while keeping memory bounded, but most current methods leave alerting threshold selection as a post-hoc tuning problem poorly suited to production. Operators need alerting behaviour specifiable before deployment using inputs such as false-negative cost, false-positive cost, and alerting budget. This paper presents CALIBURN, a five-component streaming alerting pipeline composed of a truncated Bayesian online change-point detector, an isotonic calibration layer mapping the change-point posterior to an empirical conditional attack probability, a cost-sensitive decision threshold derived from operator-specified misclassification costs, a Conformal Risk Control wrapper that converts an alert-budget specification into a within-window valid threshold under exchangeability, and a multi-window burn-rate alerting layer adapted from Site Reliability Engineering practice. Rather than claiming uniform dominance, we present CALIBURN as a regime-sensitivity study, evaluating the pipeline across three attack-prevalence regimes: LITNET-2020 at 5.2 percent, CICIDS2017 at 22.06 percent, and UNSW-NB15 at 64 percent. In the rare-attack regime, CALIBURN achieves AUC-PR 0.943 on LITNET-2020, outperforming the best streaming baseline by 2.21x and the best batch reference by 4.12x; isotonic calibration reduces Brier score by 30 percent. In the moderate-prevalence regime, CALIBURN remains the strongest streaming method on CICIDS2017 but is exceeded by batch density methods. In the high-prevalence regime, all streaming methods approach the prevalence floor. We further identify two distinct CRC-collapse mechanisms driving the alert rule to degeneracy at small alpha, treating both as operational guidance for practitioners.