The rapid advancement of generative adversarial networks and diffusion models has led to AI-generated images approaching photorealistic quality, posing a grave threat to digital media authenticity and societal trust. Existing detection methods suffer from inadequate generalisation capabilities due to difficulties in capturing multiscale artefacts. This study proposes a discriminative network incorporating residual attention mechanisms. By embedding parallel channels and spatial attention modules within a ResNet-50 backbone, the model achieves adaptive focus on critical regions of synthetic traces. Testing on a dataset comprising 140,000 images demonstrates that this approach achieves an accuracy of 97.85% and an AUC value of 0.9968, significantly outperforming mainstream baseline models. Ablation experiments validate the necessity of each component, providing an effective solution for identifying AI-generated content.
