MFI-Net: multi-level feature invertible network image concealment technique.

Area of Science:

• Computer Vision
• Deep Learning
• Information Security

Background:

• Deep learning and invertible networks offer secure image hiding, but existing methods struggle with precise region selection and exploiting low-level features.
• Limitations include reduced generation quality, overfitting, and poor generalization.

Purpose of the Study:

• To propose a novel image hiding method, MFI-Net, that overcomes limitations of existing techniques.
• To enhance the quality and security of hidden information in images.

Main Methods:

• Developed MFI-Net, incorporating a new upsampling convolution block (UCB) and a residual dense block with PReLU activation.
• Introduced a novel frequency domain loss (FDL) to guide secret information concealment into suitable image regions.
• Utilized multi-level image features (low-level and high-level).

Main Results:

• MFI-Net demonstrated superior performance over state-of-the-art methods on DIV2K, COCO, and ImageNet datasets.
• Achieved superior image quality metrics and successful application to digital collection images.
• The method effectively utilizes multi-level features and the FDL enhances concealment precision.

Conclusions:

• MFI-Net represents a significant advancement in secure image hiding using invertible networks.
• The proposed UCB, PReLU-enhanced residual dense block, and FDL contribute to improved quality, security, and generalization.
• MFI-Net offers a robust solution for concealing information within images effectively.