Scholarlink

<p>Digital images circulate widely across journalism, law and social media, yet are increasingly vulnerable to sophisticated manipulations: copy–move, splicing, inpainting, deepfakes, JPEG recompression and EXIF metadata tampering. Most detectors focus on a single manipulation family and lack unified reasoning across visual and metadata cues. We propose a compact multi-branch system combining Vision Transformer (ViT) classifiers for structural forgeries, EfficientNet-B0 for deepfakes, LightGBM for metadata anomalies and YOLOv8-Seg for pixel-level localization. The framework issues global tamper likelihoods and segmentation masks while remaining feasible on modest GPU hardware. Experiments on CASIA v2.0, deepfake sets and a curated EXIF corpus show competitive accuracy (up to 88–93%), high precision (up to 98%) and localization IoU above 0.90.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Unified framework for image forgery detection and localization using multi-branch deep models

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Vitthal B. Kamble,
  • Nilesh J. Uke

摘要

Digital images circulate widely across journalism, law and social media, yet are increasingly vulnerable to sophisticated manipulations: copy–move, splicing, inpainting, deepfakes, JPEG recompression and EXIF metadata tampering. Most detectors focus on a single manipulation family and lack unified reasoning across visual and metadata cues. We propose a compact multi-branch system combining Vision Transformer (ViT) classifiers for structural forgeries, EfficientNet-B0 for deepfakes, LightGBM for metadata anomalies and YOLOv8-Seg for pixel-level localization. The framework issues global tamper likelihoods and segmentation masks while remaining feasible on modest GPU hardware. Experiments on CASIA v2.0, deepfake sets and a curated EXIF corpus show competitive accuracy (up to 88–93%), high precision (up to 98%) and localization IoU above 0.90.