{"title":"MFI-Net: multi-level feature invertible network image concealment technique.","authors":"Dapeng Cheng, Minghui Zhu, Bo Yang, Xiaolian Gao, Wanting Jing, Yanyan Mao, Feng Zhao","doi":"10.7717/peerj-cs.2668","DOIUrl":null,"url":null,"abstract":"<p><p>The utilization of deep learning and invertible networks for image hiding has been proven effective and secure. These methods can conceal large amounts of information while maintaining high image quality and security. However, existing methods often lack precision in selecting the hidden regions and primarily rely on residual structures. They also fail to fully exploit low-level features, such as edges and textures. These issues lead to reduced quality in model generation results, a heightened risk of network overfitting, and diminished generalization capability. In this article, we propose a novel image hiding method based on invertible networks, called MFI-Net. The method introduces a new upsampling convolution block (UCB) and combines it with a residual dense block that employs the parametric rectified linear unit (PReLU) activation function, effectively utilizing multi-level information (low-level and high-level features) of the image. Additionally, a novel frequency domain loss (FDL) is introduced, which constrains the secret information to be hidden in regions of the cover image that are more suitable for concealing the data. Extensive experiments on the DIV2K, COCO, and ImageNet datasets demonstrate that MFI-Net consistently outperforms state-of-the-art methods, achieving superior image quality metrics. Furthermore, we apply the proposed method to digital collection images, achieving significant success.</p>","PeriodicalId":54224,"journal":{"name":"PeerJ Computer Science","volume":"11 ","pages":"e2668"},"PeriodicalIF":3.5000,"publicationDate":"2025-02-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11888862/pdf/","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"PeerJ Computer Science","FirstCategoryId":"94","ListUrlMain":"https://doi.org/10.7717/peerj-cs.2668","RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"2025/1/1 0:00:00","PubModel":"eCollection","JCR":"Q2","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0
Abstract
The utilization of deep learning and invertible networks for image hiding has been proven effective and secure. These methods can conceal large amounts of information while maintaining high image quality and security. However, existing methods often lack precision in selecting the hidden regions and primarily rely on residual structures. They also fail to fully exploit low-level features, such as edges and textures. These issues lead to reduced quality in model generation results, a heightened risk of network overfitting, and diminished generalization capability. In this article, we propose a novel image hiding method based on invertible networks, called MFI-Net. The method introduces a new upsampling convolution block (UCB) and combines it with a residual dense block that employs the parametric rectified linear unit (PReLU) activation function, effectively utilizing multi-level information (low-level and high-level features) of the image. Additionally, a novel frequency domain loss (FDL) is introduced, which constrains the secret information to be hidden in regions of the cover image that are more suitable for concealing the data. Extensive experiments on the DIV2K, COCO, and ImageNet datasets demonstrate that MFI-Net consistently outperforms state-of-the-art methods, achieving superior image quality metrics. Furthermore, we apply the proposed method to digital collection images, achieving significant success.
期刊介绍:
PeerJ Computer Science is the new open access journal covering all subject areas in computer science, with the backing of a prestigious advisory board and more than 300 academic editors.