基於密集連接卷積神經網路之JPEG影像隱寫分析;Steganalysis in JPEG Images based on Densely Connected Convolutional Neural Network

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/81326

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/81326

Title:	基於密集連接卷積神經網路之JPEG影像隱寫分析;Steganalysis in JPEG Images based on Densely Connected Convolutional Neural Network
Authors:	胡家銘;Hu, Jia-Ming
Contributors:	資訊工程學系
Keywords:	隱寫分析;隱寫術;深度學習;卷積神經網路;Steganalysis;Steganography;Deep Learning;Convolutional Neural Networks.
Date:	2019-08-20
Issue Date:	2019-09-03 15:44:41 (UTC+8)
Publisher:	國立中央大學
Abstract:	隱寫術(Steganography)是將大量資料隱藏於如影像、視訊等載體之技術，而隱寫分析(Steganalysis)為隱寫術的對抗技術，用於判斷載體中是否隱藏額外資訊。本研究藉由深度學習(Deep Learning)，訓練一個由密集連結的卷積神經網路(Convolutional Neural Networks，CNN)模塊為主體，設計一個針對JPEG影像隱寫術的全新隱寫分析架構。本架構結合Inception Net、ResNet與DenseNet的設計巧思與優勢，包括Inception Net中結合多種不同尺度卷積核以增加對尺度的適應性，ResNet對於特徵擷取的高重複使用率且冗餘程度低，而DenseNet可創造新特徵但冗餘程度高等特性，設計出效果更好的隱寫分析架構。本研究的模型訓練不需要人工預處理，可自動學習有效特徵。在特徵提取的部分，本機制不使用池化操作以避免抑制隱寫訊號，偵測則以三種隱寫術為對象。此外，我們設計所謂「多模型決策方法」，結合多個單一模型一起檢測，並以原先的個別單一訓練模型做為比較對象。實驗結果顯示我們的模型和方法，幾乎都超越了比較的目標，甚至是目前最先進的隱寫分析模型-SRNet。;Steganography is a technique for hiding large amounts of data in carriers such as video and video, and Steganalysis is a technique to determine whether additional information is hidden in the carrier. In our study, Deep Learning is used to train a densely connected Convolutional Neural Networks (CNN) module to design a new steganalysis architecture for JPEG image steganography. This architecture combines the design ingenuity and advantages of Inception Net, ResNet, and DenseNet.Including the inclusion of multiple scale convolution kernels in Inception Net to increase scalability, and ResNet′s high reuse and low redundancy for feature extraction. DenseNet can create new features but high levels of redundancy to design a better steganalysis architecture. The model training of our study does not require manual preprocessing and can automatically learn effective features. In the feature extraction part, the module does not use the pooling operation to avoid suppressing the steganographic signal, and the detection is based on three steganography. In addition, we design a so-called "multi-model decision method" that combines multiple single models to detect together and compares with the original individual training models. The experimental results show that our models and methods almost surpass the goal of comparison, even the most advanced steganalysis model - SRNet.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	253	View/Open

社群 sharing

Loading...