Many real-world problems can be solved using deep learning, including image processing and speech recognition. However, deep learning and neural networks are usually black boxes, making them vulnerable to threats from abnormal examples, such as adversarial and backdoor examples crafted by attackers with malicious intentions and, even worse, unintentionally mislabeled examples. The detection of such abnormal examples is therefore urgent and important. Various approaches have been used to detect specific types of abnormal examples, but all have limitations. Despite these limitations, this problem remains a significant one. To make AdvGAN more suitable for adversarial training, we propose updating its loss terms and testing it against various attacks. In contrast to generative models with no adversarial training, generative models with adversarial training are better at boosting adversarial robustness. Experimental results demonstrate a high true-positive rate (98.11%) and a low false-positive rate (3%).