Network intrusion detection must reduce false alarms while catching attacks, yet data privacy prevents pooling traffic across sites and models are heterogeneous. We present a privacy-preserving, score-level ensemble that fuses only class probabilities from multiple NIDS. For each class, we define utility as average precision and compute exact Shapley values over model coalitions to obtain a model×class weight matrix. The weighted probabilities yield a global decision and can be updated in a sliding window without sharing raw data or parameters. On a public dataset our method outperforms Equal and Static weighting. The approach amplifies specialization, suppresses redundancy, and aligns with operational constraints.