MVSS-Net: Multi-View Multi-Scale Supervised Networks for Image Manipulation Detection.

Area of Science:

• Computer Vision
• Digital Forensics
• Machine Learning

Background:

• Image manipulation poses challenges for media forensics, leading to potential misinterpretation.
• Current deep learning methods struggle with generalization on unseen data and lack specificity for authentic images.

Purpose of the Study:

• To develop a generalized deep neural network for detecting various image manipulations.
• To improve the robustness and specificity of image forensics techniques.

Main Methods:

• Proposed a multi-view feature learning approach to exploit tampering boundary artifacts and image noise.
• Trained the network (MVSS-Net++) with multi-scale supervision (pixel, edge, image) for enhanced learning from authentic images.

Main Results:

• MVSS-Net++ demonstrated superior performance in both within-dataset and cross-dataset experiments.
• The enhanced network showed improved robustness against common image degradations like JPEG compression, Gaussian blur, and re-capturing.

Conclusions:

• Multi-view feature learning offers a promising direction for creating generalizable and robust image manipulation detection systems.
• MVSS-Net++ provides a significant advancement in media forensics, addressing limitations of current deep learning approaches.