Detecting DDPM-Manipulated Medical Images Using Contrastive Learning-based Pre-Training

Abstract

Deep learning models can detect AI-generated images with high accuracy; however, classifying locally altered fake images is more difficult due to the low manipulation rate and the preservation of original patterns. This study proposes a novel CNN architecture for detecting images manipulated with tumor in-painting based on DDPM, which are difficult to distinguish by both the human eye and neural networks. The proposed architecture was trained with a two-stage training strategy that enhances performance, and a contrastive learning approach was used in the pre-training process. The default and two-stage training performances of the presented model were compared with pre-trained neural networks such as ResNet and MobileNet under the same conditions. In addition, comprehensive performance comparisons were carried out with existing deepfake production and manipulation detection methods in the literature. As a result of the evaluations, the proposed model demonstrated competitive performance, achieving an F1 score exceeding 99% under the evaluated DDPM-based manipulation setting. © 2026 IEEE.