Unsupervised Feature Learning via Convolutional Autoencoders for Cross-Manuscript Comparison in Historical Cryptanalysis

Abstract

The study of historical enciphered manuscripts is fundamental to understanding our cultural heritage, yet a vast corpus of these archives remains inaccessible due to the complexity of ancient cryptographic systems. Traditional analysis relies heavily on manual expertise, a process that is labor-intensive and difficult to scale across the immense volume of unstudied documents. This paper proposes a novel, fully unsupervised framework for the automated comparative analysis of images of historical ciphers. Our approach leverages Convolutional Autoencoders (CAE) to learn intrinsic morphological features directly from manuscript images, bypassing the need for labeled datasets or prior knowledge of the cipher keys. By projecting symbols into a high-dimensional latent space, the system generates a “similarity fingerprint” for each manuscript, enabling a quantitative comparison of diverse documents. Experimental results demonstrate that this method effectively identifies relationships between ciphers, grouping them by cryptographic tradition. This framework provides historians with a powerful computational tool to detect shared lineages and map the evolution of secret communication across history.