Optimization of Manipulation Detection in Colonoscopy Image Generation Using StyleGAN2-ADA

We propose KY-MobileNetV2, a lightweight detector for manipulated colonoscopy images. The model targets subtle GAN-based fingerprints by amplifying high-frequency residuals with Spatial Rich Model (SRM) and High-Pass Filters (HPF), followed by per-class top-K Multiple-Instance Learning (MIL) pooling. Training was binary (Real vs. Manipulation) within Normal and Polyp subsets. A total of 13,373 images (3,313 Normal; 10,060 Polyp) were collected from Kvasir, GastroVision, HyperKvasir, and PolypGen. External validation used the Konyang University Hospital dataset (15,675 images; KYU IRB 2025-09-005). All images were standardized to 512×512. Manipulated data were synthesized using StyleGAN2-ADA with class-separated training to generate Manipulation-Normal (MN) and Manipulation-Polyp (MP). Generated images showed acceptable quality: FID 73-84, Precision 0.66-0.82, Recall 0.31-0.46, Density 0.71-1.10, and Coverage 0.88-0.92. KY-MobileNetV2 achieved 0.93-0.96 accuracy, outperforming MobileNetV1 (0.88-0.89) and MobileNetV2 (0.84-0.94). External testing yielded 0.59-0.67 accuracy. These findings indicate that a lightweight architecture can effectively detect StyleGAN2-ADA manipulations in colonoscopy images and may support deployment in resource-limited clinical environments.