Multi-resolution transfer learning for tampered image classification using SE-enhanced fused-MBConv and optimized CNN...

Area of Science:

• Computer Vision
• Digital Forensics
• Machine Learning

Background:

• Digital image tampering is prevalent, necessitating robust detection systems for forensics, journalism, and cybersecurity.
• Existing methods, including handcrafted and deep learning approaches, often struggle with subtle artifacts and generalization across diverse image manipulations.

Purpose of the Study:

• To develop a highly accurate and generalizable tampered image classification model.
• To overcome the limitations of traditional methods and existing deep learning models in detecting image manipulations.

Main Methods:

• Utilized transfer learning with the EfficientNetV2B0 backbone for feature extraction.
• Integrated a lightweight, regularized Convolutional Neural Network (CNN) classification head.
• Optimized the model using Focal Loss to handle class imbalance and incorporated compound scaling, fused MBConv layers, and Squeeze-and-Excitation (SE) attention for enhanced robustness.

Main Results:

• Achieved exceptional performance on benchmark datasets (CASIA v1, Columbia, MICC-F2000, Defacto), with Area Under the Curve (AUC) scores up to 1.0000 and F1-scores up to 0.9997.
• Outperformed 42 state-of-the-art models in accuracy, precision, recall, and generalization, especially on high-resolution and compressed images.
• Demonstrated superior performance compared to models like IML-ViT, MVSS-Net++, ConvNeXtFF, and DRRU-Net.

Conclusions:

• The proposed model offers practical effectiveness and forensic reliability for digital image tampering detection.
• The system's robustness and generalization capabilities make it suitable for real-world applications in critical domains.
• The integration of EfficientNetV2B0, attention mechanisms, and Focal Loss provides a powerful approach to image forensics.