Multi-resolution transfer learning for tampered image classification using SE-enhanced fused-MBConv and optimized CNN heads

The widespread use of digital image tampering has created a strong need for accurate and generalizable detection systems, especially in domains like forensics, journalism, and cybersecurity. Traditional handcrafted methods often fail to capture subtle manipulation artifacts, and many deep learning approaches lack generalization across diverse image sources and manipulation techniques. To address these limitations, we propose a tampered image classification model based on transfer learning using EfficientNetV2B0. This backbone is combined with a lightweight, regularized CNN classification head and optimized using Focal Loss to address class imbalance. The architecture integrates compound scaling, fused MBConv layers, and squeeze-and-excitation (SE) attention to improve feature representation and robustness. We evaluate the model on four benchmark datasets-CASIA v1, Columbia, MICC-F2000, and Defacto (Splicing)-and achieve exceptional performance, with AUC scores up to 1.0000 and F1-scores up to 0.9997. Comparisons with 42 state-of-the-art models, including IML-ViT, MVSS-Net++, ConvNeXtFF, and DRRU-Net, show our method consistently outperforms existing approaches in accuracy, precision, recall, and generalization, particularly on high-resolution and compressed images. These results demonstrate the practical effectiveness and forensic reliability of the proposed system.