Publication Details

Abstract

Digital images and videos are among the most widely used media for communication, information sharing, and digital evidence. The growing availability of advanced image editing software and Artificial Intelligence (AI) generation tools has made manipulated multimedia content increasingly common and increasingly difficult to detect visually. This paper presents a web-based system for recognizing digitally altered images and videos by combining a fine-tuned MobileNetV2 deep learning classifier with complementary digital forensic techniques, namely Error Level Analysis (ELA), metadata inspection, edge detection, noise analysis, sharpness evaluation, and color inconsistency analysis. When the trained model is unavailable, the system automatically falls back to a weighted forensic-scoring pipeline so that classification is never interrupted. Videos are processed by extracting representative frames, analyzing each independently, and aggregating the frame-level scores into an overall verdict. The system is implemented using the Flask framework with TensorFlow/Keras for model inference, OpenCV for image and video processing, and SQLite for user and analysis-record management, with secure authentication for both end users and administrators. Experimental evaluation shows that the MobileNetV2-based classifier achieves an accuracy of 97.80%, precision of 96.73%, recall of 98.68%, and an F1-score of 97.71%, confirming that combining learned visual features with handcrafted forensic evidence produces a reliable, transparent, and practical solution for multimedia authenticity verification.