Multi-View GAN-Based Intrusion Detection with Temporal Reconstruction and Class-Aware Feature Balancing

Abstract

Intrusion Detection Systems (IDS) deployed in modern networks face persistent challenges arising from severe class imbalance, latent temporal dependencies in traffic behavior, and the limitations of purely discriminative learning models. Existing deep learning-based IDS approaches typically address these issues independently, relying on static feature modeling, heuristic data balancing, or stand-alone temporal anomaly detection. This paper presents Multi-View Generative Adversarial Network (MV-GAN-IDS), a unified and dataset-agnostic intrusion detection framework built upon behavioral reconstruction and class-aware generative balancing within a multi-view learning architecture. The proposed design employs a reconstruction-driven Behavioral Reconstruction Module (BRM) to capture short-range temporal inconsistencies in network flows and a Class-Aware Generative Balancing Module (CGBM) to selectively enhance minority attack representation in the tabular feature space. These complementary signals are fused through a lightweight Transformer-based classifier that jointly models static flow attributes and behavioral deviation information. Extensive evaluations on NSL-KDD and CSE-CIC-IDS2018 demonstrate consistent improvements in minority-class detection and macro-F1 performance, while maintaining strong overall accuracy. The results confirm that modeling temporal behavior alongside class-aware generative learning enhances detection reliability under class imbalance.