REFORGE A Robust Ensemble for Image Forgery Detection and Localization in Social Network Images

Digital image manipulation on social networks requires robust forensic tools for detection and precise localization. This work presents REFORGE, an ensemble framework built on five deep learning models trained in the backend: two classifiers (F3‑Net with Swin‑Base transformer and a ResNet‑50 feature pyramid fusion network) and three segmentation models (CATNet with DCT artifacts, TransUNet‑Swin, and YOLOv8‑seg). For deployment, the three best‑performing models are selected to form the frontend pipeline. An input image is first classified by F3‑Net as authentic or tampered. If tampering is detected, two complementary segmentation branches activate: TransUNet‑Swin generates a dense pixel‑wise forgery mask, while YOLOv8‑seg produces instance‑level polygons and bounding boxes for manipulated regions. This multi‑stage design captures copy‑move, splicing, and other forgery types. Implemented in PyTorch and deployed as a local Flask web application, REFORGE ensures user privacy by processing images entirely on‑device. Forensic analysts receive interpretable visual outputs including heatmaps, binary masks, and confidence‑annotated contours, enabling transparent and reliable examination without external data transmission.

Keywords: Image forgery detection, forgery localization, deep ensemble, transformer segmentation, social network forensics, copy‑move detection, splicing detection, YOLOv8‑seg.