Copy-Move Forgery Detection (CMFD) Using Deep Learning for Image and Video Forensics.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• The proliferation of sophisticated fake images necessitates advanced detection algorithms.
• Copy-move image forgery, a common manipulation, poses challenges for traditional methods.
• Deep learning offers improved performance but struggles with generalization and hyperparameter tuning.

Purpose of the Study:

• To develop and compare two deep learning approaches for detecting copy-move image forgeries.
• To evaluate the impact of network depth on model performance metrics (precision, recall, F1 score).
• To address generalization issues using diverse datasets and compare model efficiency.

Main Methods:

• Proposed two deep learning models: a custom architecture and a transfer learning model (VGG-16).
• Analyzed network depth's effect on precision, recall, and F1 score.
• Tested generalization across eight open-access image datasets.

Main Results:

• The VGG-16 transfer learning model achieved approximately 10% higher performance metrics than the custom architecture.
• Both models were evaluated on generalization capabilities across multiple datasets.
• Inference time for the transfer learning model was roughly double that of the custom model.

Conclusions:

• Transfer learning models, like VGG-16, offer superior performance for copy-move forgery detection.
• Custom architectures may provide faster inference times, presenting a trade-off between accuracy and speed.
• Further research can optimize deep learning models for robust and efficient fake image recognition.