Skip-VFL: Mitigating Extreme Class Imbalance in Feature-Heterogeneous Vertical Federated Learning for Intrusion Detection

Distributed network environments increasingly rely on Vertical Federated Learning (VFL) to address the growing heterogeneity of device features, which conventional Federated Learning is ill-equipped to handle. While VFL facilitates collab- oration across diverse network entities in detecting anomalies, current models often remain impractical for long-tailed datasets, especially in network security domain, as they fail to account for the inherent class imbalances that bias models toward majority classes. Therefore, We propose Skip-VFL, a novel tiered framework for robust anomaly detection in heterogeneous networks. Skip- VFL organizes data into three categories: shared core fea- tures (Fcore), overlapping features (Foverlap), and unique features (Funique), bridging the gap between resource-constrained IoT devices and high-capacity servers. To address class imbalances, Skip-VFL introduces a two-stage training approach: Latent Space Alignment, which clusters feature representations, and Weighted Classification using Inverse Class Frequency (ICF) to increase sensitivity to rare anomalies. A key innovation is the Skip-Layer Mechanism, which preserves feature privacy by withholding tier-specific adapters (ϕk ) from the central server while enabling global model synchronization. Experiments on NSL-KDD, UNSW-NB15, and ToN-IoT datasets show that Skip-VFL outperforms baseline methods, achieving 88.4% recall on NSL-KDD in extreme 1% imbal- ance scenarios. Unlike peers, which struggles with instability, Skip-VFL maintains consistent performance, proving it to be a reliable, privacy-preserving solution for anomaly detection in decentralized networks.