基於分布圖位移及預測差異擴張之可逆式加密音訊隱寫術

Abstract

本篇論文針對加密音訊提出一個完全可逆並具有高資訊容納量的隱寫術演算法。沿用先前加密音訊隱寫術的演算法結構：內容擁有者將音訊分成兩個組別，並使用負責加密的樣本群去預測負責嵌入資訊的樣本群，將負責嵌入資訊的樣本群改寫為預測的誤差，再透過部分加密，使訊息嵌入者能在加密的音訊中直接嵌入資料；訊息嵌入者則依照資訊大小決定使用分布圖位移或是預測差異擴張等不同的嵌入演算法；接收者可以根據實際的應用目的，彈性地決定優先解密，或是優先取出嵌入資訊，並能在取出訊息後完整地還原回原始資訊。透過引入非因果線性預測模型及分布圖位移、預測差異擴張的資訊嵌入方法，使此算法達到先前加密音訊隱寫術無法達到的完全可逆性並且擁有高資訊容納量。在實驗結果及討論中，我們發現了非因果線性預測模型的參數選擇除了要考慮平均平方誤差之外，也需要考慮峰態造成的影響；我們所提出的方法的預測方式傾向犧牲部分的不可察覺性達到更好的資訊負載量，能在較複雜的音訊中有相對好的表現。

This work presents a fully reversible data hiding algorithm for encrypted audios with flexible capacity. In the first phase, a content owner exploits the noncausal linear prediction model to divide audio sequences into two groups: for embedding and encryption. Then the content owner encrypts the group for encryption. Through this, the data hider can directly embed messages without knowing the original content. After receiving the encrypted audio, a data hider can choose two kinds of embedding methods: histogram shifting and prediction-error expansion, depending on the message length. Therefore, this algorithm reaches a flexible capacity without losing reversibility. At the receiver side, a receiver can recover the original content after extraction of the embedding message. Moreover, the order of decryption and data extraction can be changed depending on the practical applications. After the experiment, we found that the chosen coefficient of the noncausal linear prediction model depended on not only quadratic mean but also kurtosis. Furthermore, our scheme changes some imperceptibility for better capacity. Therefore, our scheme has a better performance when input audios become more complex.