Network security intrusion detection is becoming exponentially important with the acceleration of cyber attacks due to expanding internet. In this work, we introduce a new framework integrating Conditional Generative Adversarial Networks (CGANs) and Knowledge Distillation to enhance detection efficiency while keeping the model efficient on the UNSW-NB15 dataset. The CGAN is utilized to create realistic synthetic samples that strengthen the diversity and balance of the training data, in order to counter the class imbalance problems inherent in intrusion detection datasets. A complex and large-scale Teacher model is initially trained to acquire complex data patterns before being utilized to transfer knowledge to a compact Student model through Knowledge Distillation. Experimental results demonstrate that although the Teacher model has high accuracy (90.12%) and an F1 score (0.9227), the much smaller Student model (79.13% size reduction) has competitive performance with 89.03% accuracy and an F1 score of 0.9131. Moreover, the Student model has a lower false positive rate (1.70%) than the Teacher model (2.61%), which indicates its suitability for real-time and resource-limited environments. Our method optimally trades detection performance and deployment efficiency and is therefore well suited for real-world cybersecurity applications.