CDAF: A Contrastive Distribution Alignment Framework for Generative Augmentation in Imbalanced IIoT Intrusion Detection

In industrial Internet of Things (IIoT) intrusion detection, severe class imbalance and inter-class feature overlap often degrade the performance of traditional classifiers in recognizing minority classes. This paper proposes a Contrastive Distribution Alignment Framework (CDAF), which aims to achieve high-quality generative augmentation through feature-space alignment constraints and a curriculum-based training strategy. CDAF learns a structured latent space conditioned on the target class, enabling the generator to produce more representative and class-consistent synthetic samples. The framework further constructs “generated–target” and “generated–non-target” sample pairings, introducing a contrastive feature alignment regularization term that minimizes the distributional distance between generated samples and target-class features, while simultaneously enlarging their margin from non-target classes. Furthermore, we design a stage-wise weight transfer mechanism termed ECLG (Enhanced-Contrast Late-Generative), which emphasizes inter-class discrimination during the early training phase and focuses on generation quality in the later phase, thereby enabling dynamic optimization of the generative distribution. Empirically, we show that CDAF’s contrastive enhancement loss effectively reduces feature overlap under inter-class distance constraints, and the overall framework significantly improves minority-class detection on three benchmark IIoT datasets (X-IIoTID, WUSTL-IIoT-2021, and ToN IoT).