DEEP LEARNING-BASED CLASSIFICATION OF AUTHENTIC AND TAMPERED IMAGES WITH TAMPERING LOCALIZATION VIA IMAGE PROCESSING

Abstract

Digital photographs underpin evidence in journalism, law, and medicine, yet skillfully forged images can still evade casual inspection. This paper introduces an InceptionV3-based network that simultaneously classifies images as authentic or tampered, while also identifying the manipulated regions. Experiments were conducted on the CASIA v2.0 dataset, which contains 7,492 authentic and 5,123 tampered images with 5,123 corresponding ground truth images. We use an 85% / 7.5% / 7.5% split for training, validation, and testing. All images are resized to 224 x 224 px and normalized to [0, 1] be-fore end-to-end training with Adam (initial learning rate 1 x 10⁻⁴) and categorical cross-entropy. The proposed model attains 95.31% accuracy on the test set, surpassing recent InceptionV3 baselines despite its compact architecture. Tamper localization is achieved by thresholding the network’s dual-branch saliency map, which is then overlaying binary ground-truth contours. These overlays accurately highlight boundaries, even under aggressive JPEG compression and minor edits. By sharing features between detection and localization, the network learns forensic artifacts blocking, boundary mismatches, and illumination discontinuities, without the need for auxiliary loss terms or attention modules. Code, trained weights, and evaluation scripts are made publicly available to ensure full reproducibility.