In the digital editing age, it is crucial to determine image forgeries because the authenticity and integrity of visual images are being at a higher compromise. The presented work introduces the XFMA-Net (EfficientNet with Built-in Multi-Scale Attention), the innovative architecture of deep learning that can be used to recognize different image forgeries such as splicing, copy-move, and retouching. XFMA-Net is built based on the EfficientNet-B7 architecture and consists of multi-scale feature fusion and attention mechanisms, which improves the representability of the features and increases its effectiveness as a forgery detector. The model was assessed using the CASIAv2.0 dataset, exhibiting enhanced performance relative to current methodologies. XFMA-Net attained an exceptional accuracy of 98.3%, precision of 97.8%, recall of 98.2%, and F1-score of 98%, markedly surpassing contemporary methodologies. Alongside robust quantitative outcomes, XFMA-Net prioritises model interpretability by facilitating attention-guided visualisation of manipulation-sensitive areas through Grad-CAM-based localisation. These visual elucidations underscore the areas that predominantly affect the model’s determinations, offering transparent and interpretable forensic proof of image manipulation.