Asymmetric class-conditional routing in a federated quantum mixture-of-experts improves imbalanced intrusion detection in edge industrial IoT networks

Intrusion detection systems (IDS) deployed in Edge Industrial Internet of Things (Edge-IIoT) environments face three critical and interrelated challenges: (i) extreme class imbalance, where tail-class attack categories may be outnumbered by benign traffic at ratios exceeding 700:1; (ii) competitive expert collapse in gating mechanisms relying on softmax normalization, which prevents sparse, class-conditional routing; and (iii) statistical heterogeneity across geographically distributed edge nodes with non-IID local traffic distributions. Existing federated learning solutions address these challenges in isolation, yielding suboptimal tail-class detection. We propose FedQMoE-IDS , a novel federated quantum mixture-of-experts framework that jointly resolves all three challenges through three synergistic innovations: (1) a class-balanced asymmetric focal loss with per-class focusing parameters \(γ _c\) that independently modulate head- and tail-class gradients via effective-number weighting; (2) a normalized sigmoid gating mechanism with auxiliary-loss-free dynamic bias load balancing, replacing softmax to achieve genuine sparse activation with exact zero weights and class-conditional expert routing; and (3) a selective federated aggregation strategy that partitions the parameter space into shared (FedDyn-aggregated CNN+KAN+classifier), locally retained (personalized gating networks), and periodically exchanged (variational quantum circuit experts) components. The three-stage MoE architecture cascades 1D-CNN feature extractors, variational quantum circuit (VQC) encoders operating in Hilbert space, and Kolmogorov–Arnold Network (KAN) spline experts, each governed by independent sigmoid top- K gates. We conduct experiments under a realistic non-IID federated setting with \(N{=}10\) Edge-IIoT clients partitioned via a Dirichlet distribution ( \(α {=}0.3\) ), and we compare against parameter-matched classical baselines (deeper CNN and softmax-gated MoE), personalized federated MoE (FedMoE, pFedMoE), and federated class-imbalance methods. Comprehensive experiments on UNSW-NB15, Edge-IIoTset, and CIC-IDS2017 with five-seed evaluation and Bonferroni-corrected Wilcoxon signed-rank tests demonstrate that FedQMoE-IDS achieves 99.20% accuracy and 98.40% macro F1-score on UNSW-NB15, and 99.80% accuracy on Edge-IIoTset—surpassing recent state-of-the-art methods including centralized deep learning, federated LSTM, and transformer-based IDS. All improvements are statistically significant ( \(p < 0.05\) ). The full source code and configuration files are publicly available to enable independent reproduction.