With the advancement of deep learning, intrusion detection technology has significantly improved, but it still faces two major problems: first, network attacks constantly evolve in practice, making trained models struggle to detect new unknown attacks; second, network traffic suffers from imbalanced benign and attack data. To address these challenges, the ATGCB intrusion detection model is proposed. It consists of an adaptive module and a multi-generator adversarial network. The adaptive module is an unsupervised clustering model that detects data drift—when unknown attacks are identified, the generation module activates to produce data for training new classifiers. The multi-generator adversarial network generates synthetic samples to augment minority classes and balance the dataset. Experiments on the CICIDS-2017 and CSE-CICIDS-2018 datasets show that after data balancing, The proposed method achieves a detection accuracy of 97.56% after balancing the data, and maintains 91.12% accuracy when handling drifted data. demonstrating its superior applicability in modern network intrusion detection.