A Minority Sample Generation Framework for Intrusion Detection Under Complex Data Distributions Based on Discriminative Representation Learning and Jointly‐Constrained Conditional Diffusion

In network intrusion detection, complex data distributions caused by class imbalance, class overlap, and missing data significantly limit the modeling and recognition performance for minority class samples. Existing sampling and generation methods are often designed for single problems, which can lead to bias toward the majority class in such complex scenarios and make it difficult to accurately model overlapping regions and incomplete data structures. In addition, these methods usually rely on external interpolation to handle missing data, which may introduce distribution bias and affect the realism of generated samples, thereby limiting improvements in downstream performance. To address these issues, this paper proposes a minority class generation framework that integrates discriminative representation learning with jointly‐constrained conditional diffusion modeling. In the feature learning stage, a center contrastive variational autoencoder (CC‐VAE) is constructed. It uses masked modeling and a self‐supervised reconstruction mechanism to learn robust representations for missing data. At the same time, center contrastive optimization is introduced to improve the discriminative ability of the latent space. In the generation stage, a Jointly‐Constrained Conditional Diffusion Model (JCCDM) guided by class prototypes is designed. Through joint constraints from the latent space and the original feature space, the proposed model generates semantically consistent and structurally reliable minority class samples. Experimental results on multiple real world cybersecurity datasets demonstrate that the proposed method improves both the modeling quality and recognition performance of minority classes, while enhancing overall intrusion detection performance. In addition, the proposed method achieves favorable overall performance and average rankings across different downstream classifiers, with statistically significant differences observed against several baseline methods.