改良式雙向聯想記憶類神經網路加解密之研究

Abstract

過去在加解密模型上的演算法大都著重於邏輯式的演算法，在建構加解密系統上較為複雜，且架構不具廣義性。由於已經有研究者提出模仿生物神經系統的人工類神經網路架構（Architecture of Artificial Neural Network, ANN）為基礎下的加解密模型（Back-propagation and Overstoraged Hopfiled Neural Network（OHNN））進行加解密，但都有其限制（例如：加密量的限制、解密後資料完整性的限制、穩定度的限制等限制因素）。 本論文提出應用類神經網路雙向聯想記憶（Bi-directional Association Memory, BAM）的演算方法建構加解密模型進行加解密，此方法使得加解密模型在建構上具有簡便性及廣義性，並且利用BAM的雙向狀態穩定的特性解決穩定度的限制。由於BAM主要利用區域極小值（local minima）儲存資料，且其學習規則是採用Hebbian 學習法，因此可能使網路能量區域極小值的數量超過原先儲存的資料量，而造成偽狀態（spurious states）的情況發生，使得資訊喪失資料完整性的原則。為了解決上述的問題，故配合空間變換（space transformation）的概念，得以避開偽狀態的影響並且增加加密量、確保解密後資料完整性的原則、降低解密時間。再利用Shannon所提出的完美秘密（perfect secrecy）的概念量化證明本系統的安全性（security）程度。

Most algorithms developed for encryption and decryption were concentrated on logic analysis. But it is complex for system construction and difficult to apply wide-spread. Recently, even though biomimetic-based architecture of artificial neural network was proposed to improve reliability and performance of encryption methods such as back-propagation and overstoraged Hopfield Neural Network were developed to fulfill this expectation. But the limitations of encryption capacity, complexity and data completeness after decryption, reliability are still needed to overcome. This paper proposed a new algorithm to improve reliability and convenience of encryption and decryption with reformed Bi-directional Association Memory (BAM) model to reduce spurious states and data separation caused by former local minima information analysis based on Hebbian learning rule. The space transformation was used to escape crosstalk and noise vector caused by spurious states to keep the completeness of processed information in addition to enhance its security. MATLAB simulation model was used to testify the performance of BAM cryptosystem. The experimental results showed that the security of this proposed system has been improved by Shannon’s perfect secrecy conception.