A multi-model feature fusion based transfer learning with heuristic search for copy-move video forgery detection

Area of Science:

• Computer Vision
• Digital Forensics
• Machine Learning

Background:

• Digital image and video data are ubiquitous, making them targets for manipulation.
• Copy-move forgery, a common video tampering technique, involves duplicating and reinserting image segments.
• Existing deep learning methods for forgery detection often require extensive training data and hyperparameter tuning.

Purpose of the Study:

• To develop an effective model for detecting and classifying copy-move forgery in video content.
• To enhance the accuracy and robustness of video forgery detection systems.
• To address the limitations of current deep learning approaches in terms of data dependency.

Main Methods:

• The proposed Enhancing Copy-Move Video Forgery Detection through Fusion-Based Transfer Learning Models with the Tasmanian Devil Optimizer (ECMVFD-FTLTDO) model processes video frames.
• Noise reduction is performed using a modified wiener filter (MWF).
• A fusion-based transfer learning (TL) approach combines ResNet50, MobileNetV3, and EfficientNetB7 models for feature extraction, followed by an Elman recurrent neural network (ERNN) classifier optimized by the Tasmanian devil optimizer (TDO).

Main Results:

• The ECMVFD-FTLTDO model demonstrated superior performance in detecting copy-move forgery.
• The model achieved high accuracy rates of 95.26% on the GRIP dataset and 92.67% on the VTD dataset.
• Comparative analyses showed significant improvements over existing forgery detection techniques.

Conclusions:

• The ECMVFD-FTLTDO model offers a robust and accurate solution for copy-move video forgery detection.
• The fusion of multiple transfer learning models and TDO-optimized ERNN effectively captures spatial features and improves detection capabilities.
• This approach represents a significant advancement in digital forensics and video security.