The article investigates the problem of critical class imbalance in intrusion detection systems (IDS) for Internet of Things (IoT) networks. A comparative study of data augmentation methods was conducted, evaluating five generative adversarial network (GAN) architectures (CopulaGAN, CTGAN, CTAB-GAN+ and modified versions of MCWGAN-GP and TMG-GAN) against traditional approaches (SMOTE, random oversampling). The study shows that data augmentation (as a data preprocessing stage) enables the restoration of the LightGBM classifier’s performance in critical imbalance scenarios, increasing the F1-macro score from 0.03 to 0.81.