<![CDATA[This document introduces a combined framework for validating digital images in forensic contexts by merging Error Level Analysis (ELA) with Convolutional Neural Networks (CNN). The innovation of this research resides in the direct integration of a conventional explainable forensic method alongside a datadriven deep learning approach to ensure both clarity and enhanced detection efficacy. ELA serves to identify JPEG compression irregularities as forensic indicators, whereas CNN is employed to extract significant hierarchical features for robust image categorization. Trials were performed on the CASIA v2.0 dataset, which comprises 10,002 authentic and altered images. The suggested two-stream architecture concurrently processes original images and ELA-generated maps, facilitating synergistic feature acquisition. The hybrid model secures an accuracy rate of 74.32%, illustrating a 7.2% enhancement over isolated ELA. Furthermore, the framework diminishes the false positive rate from 50.2% to 34.8% while maintaining high sensitivity (0.84) in identifying altered regions. From a machine learning angle, this research illustrates how manually crafted forensic attributes can boost CNN capabilities when merged at the input stage. From an image processing viewpoint, it confirms ELA as a potent preprocessing strategy for directing deep feature extraction. The proposed framework provides an equilibrium between precision and forensic transparency, making it ideal for real-world digital forensic practices, including application in environments with limited resources.]]>
Copyrights © 2025
