<p>Although deep learning-based forged image detection methods can improve detection accuracy through end-to-end feature learning, they face core challenges such as feature drift caused by the iteration of generative models, insufficient modeling of cross-resolution artifacts, and lack of robustness against post-processing. This paper proposes MIRC-FADNet: a forged image detection network for generative images with mutual information regularization constraints and frequency-domain attention, which breaks through the bottlenecks through the collaboration of three modules: First, design a deep information maximization encoder with mutual information entropy constraints, which quantifies the feature mutual information between generative and real images through deep learning feature measurement to dynamically lock stable forged fingerprints; second, introduce the MobileViTv2 module, which strengthens the artifact representation of cross-resolution pixels by utilizing the long-range dependency modeling capability of deep learning; and third, embed the sensor response model convolution layer, which enhances the recognition ability of frequency-domain distortion patterns by combining the local feature extraction advantages of deep learning. Since this deep learning model needs to process pixel features of high-resolution images and massive training data from 9 types of generative models, its complex feature calculation and parameter optimization processes rely on the parallel computing architecture of supercomputing (HPC) to achieve efficient model training and experimental verification. Experimental results show that the method achieves an average accuracy of 97.8% in cross-model detection and 93.0% in ProGAN fine-grained category detection, providing key technical support for the automated detection of generative content.</p>