Intrusion Detection Systems (IDS) based on machine learning (ML) face severe threats from adversarial attacks. A common defense method is to deploy auxiliary networks. However, these auxiliary networks, while defending against adversarial attacks, often make alterations to clean examples (examples without adversarial perturbations), which are typically negative changes. These alterations lead to a decrease in the classification accuracy of clean examples by IDS models, resulting in a high rate of false positives and false negatives. We proposed the Diverse-Task Generative Adversarial Network (DTGAN) to address this challenge. DTGAN is a training strategy for auxiliary networks based on generative adversarial networks. By analyzing the data distribution and task requirements of different types of examples, we classified the examples received by IDS and design different learning tasks for each type of example during the training of the auxiliary network. This approach results in an auxiliary network that considers the effects of handling different examples, ensuring that clean examples maintain their original data distribution after processing and avoiding any negative impact on clean examples caused by the auxiliary network. Experiment results demonstrated that DTGAN can defend against adversarial attacks while maintaining the classification accuracy of clean examples for IDS models. Through DTGAN defense, IDS models based on CNN architecture can achieve an accuracy of over 80% against various adversarial attacks. Furthermore, integrating DTGAN with adversarial training further enhances the resistance of IDS models against adversarial attacks.