影像藏密分析技術是用來偵測影像是否有藏入資料的技術,如果未能偵測到該藏密影像是否藏有秘密訊息,該秘密訊息即是安全的,但影像藏密分析技術如果只能針對單一類型藏密技術來進行分析,就有可能會偵測不到其他種類的藏密技術,本論文提出深度學習影像藏密分析,用來偵測多種影像藏密技術。本論文的影像藏密分析技術使用LSB藏密影像、DCT藏密影像及直方圖藏密影像這三種藏密影像,這些藏密影像都皆為512x512的灰階影像,針對這三種藏密影像進行偵測。因此本研究選擇兩種近年相當代表性的大型深度學習網路框架ResNet及DenseNet,將這兩種模型作為深度學習的基底框架,並導入遷移式學習的技術,載入開源的預訓練權重。雖然ResNet與DenseNet已經具有基礎的影像辨識能力,但由於這兩種框架只對ImageNet影像資料集訓練過,並沒有針對藏密影像進行訓練,因此模型還是需要透過藏密影像的再訓練,對整個模型進行權重的調整,並且再透過週期性學習率的方式,使模型藏密分析的準確率更高。實驗結果顯示本研究在影像數據量不斷增加的情況下,深度學習影像藏密分析確實有效降低誤判藏密影像的機率。未來可以透過此檢測機制來判斷影像的藏密技術是否能避開深度學習影像藏密分析的檢測,以保護影像安全。
Image steganalysis is a technique used to detect whether an image is embedded with secret data. A stego-image that cannot be detected to contain secret data implies the secret data to be safe. The current image steganalysis is only for a single type of stego-image, and cannot detect multiple images at the same time. Therefore, this paper proposed a deep learning image steganalysis to detect multiple types of stego-images. The multi-type stego-images used in this paper are LSB, DCT and histogram 512x512 grayscale images. In this study two large-scale deep learning network frameworks, namely, ResNet and DenseNet, were used as the basic framework for deep learning. The transfer learning technique is used to load open source pre-training weights. Although ResNet and DenseNet have basic image recognition capabilities, the original data of transfer learning is not pre-trained for stego-images. Therefore, the model included a training stage to adjust the weight, and to determine the learning rates through the cyclical learning rates method to increase the model's steganalytic accuracy. Results showed that increasing the amount of image dataset in this study reduced the error in the deep learning image steganalysis of the stego-images. In the future, this detection method can be used to determine whether a stego-image can avoid the detection of deep learning image steganalysis to increase the security of the stego-image.