In order to improve the detection capabilities against the attack of network application layer, the present paper proposes a novel malicious URL detection model based on Generative Adversarial Networks (GAN) and Integrated features or namely MUDGI in short. This model extracts a URL’s empirical features from the entire or part of the URL structure for grasping its structural hierarchy features holistically and identifies seven new empirical ones to improve the detection accuracy. To perform semantic feature extraction, we also develop a pooling structure based on a convolutional gated recurrent unit (GRU) via weighted fusion such that semantic features can be extracted at the word level and local features can be obtained in a detailed and comprehensive manner. Moreover, in order to enhance the detection performance of MUDGI, we further propose a fusion decision scheme based on trust and malicious bias and utilize a Generative Adversarial Network (GAN) for balancing and expanding the URL dataset. Numerical simulation over various existing and AI-generated datasets shows that the proposed MUDGI model can achieve 98.4% accuracy rate and 99.2% recall rate of malicious URL detection and have a stronger adaptive ability for zero-day attack detection than the existing detection models in the same style.