影像隱藏學之研究

Abstract

隱藏學(steganography)的起源, 最早可回溯到兩千多年以前的古希臘時代, 在希臘歷史學家 Herodotus 的史著中, 記載著幾個有名的例子: 將機密訊息刻於頭皮或蠟板下方, 讓人不察再將其傳出。隱藏學和密碼學(cryptography)最大的不同在於: 密碼學將機密訊息－明文(plaintext), 加密成密文(ciphertext), 將訊息的意義(meaning)隱藏起來; 隱藏學則是將訊息的存在性(existence)隱藏起來, 讓人完全感受不到訊息是存在的。因此, 要破解一份密文必須去猜測加密時所用的金鑰(encryption key), 將密文還原成明文; 然而, 一套訊息隱藏系統(steganographic system) , 隱藏訊息意義為何並不需要被完全理解, 只要被證明出訊息是存在的, 這套系統就沒用了。 現今, 由於電腦科技與網路通訊的高度發展, 訊息的傳遞不但快速而且便利, 而隱藏訊息的方法更是多樣化。機密訊息經加密後, 被隱藏於各式各樣的多媒體資料中, 最後透過網際網路傳送出去的是”看似無異樣” (innocuous-looking)的多媒體偽裝資料(stego-media), 以確保「通訊行為不被發現」之目的。 本論文選定數位影像為掩護媒體(cover-media), 考量秘密通訊的各種需求, 並針對影像不同的儲存格式, 分別提出三種不同的嵌入技術, 設計出安全的秘密通訊模型。首先, 針對非壓縮影像格式, 修正傳統的最低位元嵌入法(LSB embedding)的缺點, 提出一個隱蔽的高容量的多位元嵌入技術; 接著, 因應影像的壓縮標準JPEG, 提出一個在高倍率失真壓縮的情況下, 嵌入的機密訊息仍然可以正確地萃取出來的系統; 最後, 因應在網路傳遞的影像, 其主流儲存格式為JPEG與GIF, 提出一個兼容這兩類壓縮失真的嵌入技術。 在本篇論文中, 所提出的隱藏模型(steganographic model)是否夠隱蔽, 我們分別從三種不同的角度去分析: 第一, 偽裝影像(stego-image)的影像品質(fidelity), 影像中是否存在因嵌入行為所導致的不自然痕跡(artifact); 第二, 視覺攻擊法(visual attack), 因應嵌入技術, 配合人類視覺能力, 篩選出不尋常的影像; 第三, 統計攻擊法(statistical attack), 因應嵌入技術, 配合各類的統計工具, 將影像中不尋常的特徵篩選出來。 傳統的最低位元嵌入法, 已經無法通過一些最新研發出來的統計攻擊法之驗證, 針對此項缺失, 本篇論文所提出的第一項技術會依據影像中每個像素(pixel)的亮度(luminance)及與鄰近像素的對比(contrast)特性, 去估計每個像素可供嵌入的最大容量。因為所提出的是一個不同長度(variable-sized)的多位元嵌入方法, 因此目前最新的統計攻擊法仍然無法偵測出來。此外, 所產生的偽裝影像保留了原來掩護影像(cover-image)在位元平面(bit-plane)上的視覺特性, 因此若使用視覺攻擊法來分析偽裝影像是無法證明秘密通訊的行為是存在的。 接著, 分析影像壓縮標準JPEG , 我們發現JPEG壓縮的影像區塊的亮度平均值的影響是有限的, 從理論上也可以推導出JPEG壓縮所產生的誤差上限值(upper bound)。因此, 我們提出一個將訊息隱藏於亮度平均值上的嵌入技術。實驗證明, 所嵌入的機密訊息, 經過高倍率JPEG壓縮, 仍然可以完全無誤地萃取出來。實驗結果和理論推導的結論也完全吻合, 因此所提出的機密通訊模型是有效且可行的。 最後, 針對兩種在網路傳輸上最常使用的影像壓縮格式, JPEG與GIF格式, 及最新制定的影像壓縮標準JPEG 2000 (J2K), 提出一個以物件為基礎的幾何扭曲嵌入法, 來嵌入機密訊息。由於物件扭曲的程度是有限制的, 因此扭曲後的物件, 看起來仍然是非常地自然, 所以人眼完全無法察覺影像上的物件是嵌有機密訊息的。由於嵌入的過程並沒有改變物件的色彩資訊, 所以針對色彩資訊所做的各類攻擊法, 都是無效的。實驗結果也顯示, 不論是將偽裝影像存成GIF, JPEG,或J2K格式, 甚至在各種格式間轉換, 所嵌入的機密訊息, 均可以無誤地取出, 證明所提出的方法是有效的。 上述所提出的三項影像嵌入技術, 均達到在隱蔽學上的最根本需求, 以此為基礎, 使用者可依影像儲存空間大小或網路傳遞的不同頻寬需求, 選擇合適的嵌入技術, 隱藏重要的機密資訊。因此, 本篇論文所提出的嵌入技術, 是非常具有實用性的。

Steganography is an ancient art of conveying messages in a secret way that only the receiver knows the existence of the message. A classic steganographic model presented by Simmons is the prisoners' problem. In this problem, two persons in the jail plan to make an escape together. All communications between them are monitored by the warden. Therefore, they must hide the messages concerning escape plan in another innocuous-looking media. In this manner, steganography provides another layer of protection on the covert communication. The message is hidden in another media such that the transmitted data will be meaningful and innocuous-looking to everyone. Compared with cryptography attempting to conceal the content of the secret message, steganography conceals the very existence of message. On the other hand, for detecting the covert communication, the warden should analysis all messages sent between prisoners. Steganalysis is the art of detecting secret communication. If the existence of message is reveal, the purpose of steganography will be defeated. In this dissertation, the secret message is embedded in images for protecting the privacy of communication on the Internet. Since the created stego-images can be stored in various formats, we will propose three embedding techniques for covert communication under three different conditions. First, the created stego-images are stored in uncompressed formats. Second, the stego-images are compressed using JPEG standard. Finally, the format of the stego-images can be converted from JPEG to GIF, and vice versa. For analyzing the performance of the proposed steganographic models, three types of steganalysis will be considered: image fidelity, visual attack and statistical attack. Images with hidden message, called the stego-images, should be indistinguishable from their original ones, called the cover-images. Visual attack uses the ability of human eyes to clearly recognize the unusual visual patterns. Some statistical characteristics of natural images can also be used to reliably identify which images contain the hidden message. Using the least significant bit (LSB) embedding method, the hidden message can be reliably detected by using some statistical attacks. In this dissertation, a variable-sized LSB embedding method for uncompressed format will be proposed first. For investigating the maximum payload of each pixel, the number of LSBs used for message embedding depends on the local characteristics, such as luminance and contrast. Since the original bit-plane characteristics can be maintained, the hidden message is undetectable by using the visual attack. Chi-square attack is also used to show that the hidden message is statistically undetectable. Next, we will propose a variation of an information embedding method, called Quantization Index Modulation (QIM), for JPEG images. The proposed method embeds the messages to be conveyed by modifying the mean intensity value of image blocks, and the resulted stego-image can be stored in the JPEG format with a low quality setting. A theorem is also proposed to show that the error of mean intensity value of an image block caused by JPEG compression is bounded. Besides, a specific pattern caused by using QIM embedding method is also identified, and this pattern will be removed using the proposed embedding method. Experimental results and the proposed theorem show that the hidden message is error-free against the JPEG distortion under the quality setting as low as 25. Furthermore, the existence of hidden message is not only visually imperceptible but also statistically undetectable. Finally, we will present a novel idea for exactly extracting the hidden message while the stego-images are stored in various compression formats, such as GIF, JPEG and JPEG 2000 (J2K). The main idea is to embed secret data through distorting a given object and the distorted object still looks natural. Experimental results show that the hidden message can be extracted correctly even the stego-images are stored in various compression formats and rates. Furthermore, the embedded message is robust enough when the stego-image format is converted from one compression format to another one. All of the experimental results show that the proposed embedding techniques meet the fundamental requirement of steganographic system, i.e., undetectability. Therefore, selecting the most suitable technique under various transmission conditions will make the proposed steganographic model more powerful.