A lightweight AI‑generated image detection algorithm based on an enhanced MobileNetV3

This paper presents a lightweight AI‑generated image detection model based on MobileNetV3, designed to achieve high‑accuracy and low‑latency synthetic image recognition on mobile and resource‑constrained platforms. The model incorporates two core innovations: a high‑frequency detail enhancement module that extracts low‑frequency components via fixed Gaussian smoothing, computes the high‑frequency residual by differencing against the original features, and uses a learnable fusion coefficient to adaptively amplify edges and texture details; and a texture feature attention module that first captures texture edges with a Sobel operator, then employs channel and spatial attention to selectively refine and strengthen the texture channels and regions most discriminative of real versus generated images. On our dataset, the model attains a 93.14% F1‑score on the test set—an improvement of 3.49 percentage points over the original MobileNetV3—while requiring only 1.18 M parameters and delivering an inference speed of 146.9 FPS. These results demonstrate that effectively leveraging high‑frequency and texture priors can substantially enhance AI‑generated image detection performance with minimal computational overhead, providing a viable solution for rapid deployment in practical scenarios.