Error Level Analysis (ELA)-Enhanced Dual-Branch Deep Learning Model for Image Forgery Classification and Binary Mask Generation

Abstract Image forgery detection has become increasingly critical in digital era where tampered visuals can be used to manipulate evidence or spread misinformation. This paper presents an effective deep learning-based approach for classification and localization of image forgeries using Error Level Analysis (ELA) as key feature extraction technique. ELA is leveraged to amplify compression artifacts introduced during tampering, enabling enhanced visual-cues that are otherwise imperceptible in RGB images. A dual-branch CNN architecture is proposed, wherein one branch processes RGB images via a pretrained ResNet50 backbone and the other processes ELA images using a custom CNN. The fused features are used to classify images into three categories—Authentic, Copy-Move, and Spliced—and to generate a pixel-wise binary mask for forgery localization. The model was tested using ELA images generated at multiple quality levels on CASIA_v2 dataset. Results consistently showed high accuracy for Authentic and Spliced images across all quality settings. However, Copy-Move forgeries were more challenging to detect, exhibiting slightly lower precision and recall values. The highest and most balanced performance was observed at 95%, with classification accuracy of 95% and interpretable visual feedback via predicted forgery masks. This framework demonstrates strong potential for real-world image forensics applications in legal, journalistic, and cybersecurity domains.