J-SRNet: 雙域JPEG影像隱寫分析之深度學習網路設計;J-SRNet: Designing Deep Networks for Dual-Domain JPEG Steganalysis

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/98552

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98552

题名:	J-SRNet: 雙域JPEG影像隱寫分析之深度學習網路設計;J-SRNet: Designing Deep Networks for Dual-Domain JPEG Steganalysis
作者:	林語潔;Lin, Yu-Jie
贡献者:	資訊工程學系
关键词:	資料隱藏;深度學習;隱寫分析;隱寫術;JPEG
日期:	2025-08-13
上传时间:	2025-10-17 12:55:15 (UTC+8)
出版者:	國立中央大學
摘要:	PEG壓縮格式至今仍是數位影像傳輸的主流標準，大量JPEG影像儲存或傳輸於網頁、社群媒體與即時通訊軟體中。隱寫術(steganography)的目的是在多媒體載體中藏入資訊以達秘密通訊的效果，JPEG影像因此成為隱寫術中的理想載體，即利用看似正常的JPEG影像傳輸秘密資訊。另一方面，JPEG隱寫分析(steganalysis)則是從檢測的JPEG影像中判斷是否存在隱藏資訊，目前多利用深度學習技術。JPEG影像應縮將像素點利用離散餘弦轉換(DCT)產生頻率域係數並進行量化處理，這使得修改量化後的DCT係數成為JPEG隱寫術嵌入的主要方法，理論上將隱寫分析直接施於量化後DCT係數應有所助益。然而，量化後DCT係數值高度稀疏，直接輸入深度學習卷積網路容易導致梯度消失而使訓練不穩定，削弱模型對付低幅度隱寫的能力。本研究提出雙通道雙域隱寫分析架構J-SRNet，架構中包含兩條分支：空間域分支延續既有SRNet空間域偵測方法，聚焦於像素層的統計特徵；頻率域分支則針對壓縮域特性重新設計前處理流程與特徵提取子網路。具體而言，頻率域分支將DCT量化係數進行絕對值計算與裁剪，並透過位元平面（bit-plane）編碼轉換為多通道二值表示，接著使用區塊級卷積捕捉JPEG量化區塊中的特徵，整合校正模組與通道投影機制強化表徵能力，最終將兩分支所提取的特徵進行融合完成隱寫檢測任務。本研究採用BOSSBase 1.01及BOWS2資料集進行效果評估，在與目前被廣泛引用的頻率域隱寫分析開源實作比較中均取得更佳的偵測準確率，顯示J-SRNet的顯著優勢。;JPEG remains the dominant standard for digital image transmission, widely used across web pages, social media, and instant messaging platforms. As a result, JPEG images serve as ideal carriers for steganography, which embeds hidden information within multimedia content for covert communication. JPEG steganalysis aims to detect such hidden content and has increasingly adopted deep learning approaches. Since JPEG compresses images using the Discrete Cosine Transform (DCT) followed by quantization, most steganographic methods manipulate quantized DCT coefficients for data embedding. Although analyzing these coefficients directly is theoretically advantageous for steganalysis, their high sparsity poses challenges, such as gradient vanishing during training, leading to instability and poor performance, particularly against low-payload steganography. To address these issues, this thesis proposes J-SRNet, a dual-domain, dual-channel steganalysis framework that combines spatial and frequency domain processing. The spatial branch builds on existing pixel-level methods, while the frequency branch is specifically tailored to the compressed JPEG domain. It applies absolute value transformation and clipping to quantized DCT coefficients, encodes them into multi-channel binary bit-planes, and extracts features using block-level convolutions aligned with JPEG′s 8×8 quantization blocks. Calibration modules and channel projection mechanisms further enhance feature representation. Finally, features from both domains are fused to detect steganographic signals. Experiments on the BOSSBase 1.01 and BOWS2 datasets demonstrate that J-SRNet outperforms widely-used open-source frequency-domain steganalysis models, highlighting its effectiveness and robustness.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	55	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....