<p>Image forgery detection represents a fundamental task within the domain of cybersecurity. While Convolutional Neural Networks (CNNs) dominate the current landscape of image forgery detection methods, their inherent limitation lies in their focus on extracting local features, often at the expense of capturing global context. To address this limitation, this study proposes a dual-encoder architecture that integrates Mamba and ConvNeXt, enabling the comprehensive extraction of both local details and global features. We further enhance the feature representation by incorporating a ATTEN Block, which adaptively reweights the feature channels to emphasize the most salient aspects. Additionally, to mitigate the issue of missed detections caused by the complex boundaries of forged regions, edge loss is incorporated to improve the model’s accuracy in identifying forgery contours. The experiments were conducted on four publicly available image forgery detection datasets. The results show that on the CASIAv1 and Coverage datasets, the AUC and F1 scores were 97.6% and 78.1%, and 85.3% and 43.1%, respectively. On the Columbia and NIST16 datasets, the AUC and F1 metrics were 75.2% and 91.5%, and 30.8% and 77.5%, respectively. Compared to state-of-the-art methods, our approach outperformed them, showcasing superior detection accuracy and strong practical applicability.</p>