Auto-redirect paused manually.
DEEP LEARNING-BASED CLASSIFICATION OF AUTHENTIC AND TAMPERED IMAGES WITH TAMPERING LOCALIZATION VIA IMAGE PROCESSING
Abstract
Abstract
Digital photographs underpin evidence in journalism, law, andmedicine, yet skillfully forged images can still evade casualinspection. This paper introduces an InceptionV3-based network thatsimultaneously classifies images as authentic or tampered, whilealso identifying the manipulated regions. Experiments wereconducted on the CASIA v2.0 dataset, which contains 7,492authentic and 5,123 tampered images with 5,123 correspondingground 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 andnormalized to [0, 1] be-fore end-to-end training with Adam (initiallearning rate 1 x 10β4) and categorical cross-entropy. The proposedmodel attains 95.31% accuracy on the test set, surpassing recentInceptionV3 baselines despite its compact architecture. Tamperlocalization is achieved by thresholding the networkβs dual-branchsaliency map, which is then overlaying binary ground-truthcontours. These overlays accurately highlight boundaries, evenunder aggressive JPEG compression and minor edits. By sharingfeatures between detection and localization, the network learnsforensic artifacts blocking, boundary mismatches, and illuminationdiscontinuities, without the need for auxiliary loss terms or attentionmodules. Code, trained weights, and evaluation scripts are madepublicly available to ensure full reproducibility.
Digital photographs underpin evidence in journalism, law, andmedicine, yet skillfully forged images can still evade casualinspection. This paper introduces an InceptionV3-based network thatsimultaneously classifies images as authentic or tampered, whilealso identifying the manipulated regions. Experiments wereconducted on the CASIA v2.0 dataset, which contains 7,492authentic and 5,123 tampered images with 5,123 correspondingground 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 andnormalized to [0, 1] be-fore end-to-end training with Adam (initiallearning rate 1 x 10β4) and categorical cross-entropy. The proposedmodel attains 95.31% accuracy on the test set, surpassing recentInceptionV3 baselines despite its compact architecture. Tamperlocalization is achieved by thresholding the networkβs dual-branchsaliency map, which is then overlaying binary ground-truthcontours. These overlays accurately highlight boundaries, evenunder aggressive JPEG compression and minor edits. By sharingfeatures between detection and localization, the network learnsforensic artifacts blocking, boundary mismatches, and illuminationdiscontinuities, without the need for auxiliary loss terms or attentionmodules. Code, trained weights, and evaluation scripts are madepublicly available to ensure full reproducibility.