Introduction
Deciphering historical documents is a complex, multi-stage process. Before a cryptanalyst can find a secret key or decode a message, they must first identify the encipherment method used. With dozens of classical cipher types—ranging from the Vigenère to the Playfair—this initial step is often a manual, error-prone "guess-and-check" process. This paper presents an automated solution: a high-performance classifier capable of identifying 50 different classical encryption methods with state-of-the-art accuracy.
Methodology
The researchers framed cipher identification as a multi-class classification problem. To overcome the lack of large-scale "real-world" historical datasets, they developed a pipeline for infinite data generation. By using a large English corpus, they could generate random plaintexts and keys to produce a virtually limitless supply of labeled training data for each of the 50 cipher types.
The study compared several machine learning models, including Random Forests, Support Vector Machines (SVMs), and Vowpal Wabbit (VW). While SVMs struggled to scale with the massive amount of synthetic data, the VW linear classifier efficiently processed up to 1 million training examples, leading to superior performance.
Feature Engineering
A critical component of the model's success was the development of specialized features. In addition to 22 standard statistical features like the Index of Coincidence, the authors introduced three novel feature sets designed to capture the unique signatures of specific ciphers:
- REP (Repetition): Detects symbol repetitions characteristic of certain methods.
- AMSC (Amsco): Specifically targets the AMSCO cipher by analyzing grid-based permutations.
- VAR (Variant): Uses frequency profiles of columns to identify periodic ciphers.
Results and Key Insights
The best-performing model, VW-1M, achieved an accuracy of 58.49% on a benchmark set of 305 real-world ciphers from the American Cryptogram Association (ACA). This represents an 11.2% absolute improvement over the previous state-of-the-art system, BION.
The results highlighted two major takeaways:
- Scale Matters: Increasing the training data from 100,000 to 1 million examples significantly boosted the performance of the linear VW model, allowing it to outperform more complex models trained on smaller datasets.
- Domain Knowledge is Vital: As shown in the performance graph, the addition of the three novel features (REP, AMSC, VAR) provided a distinct jump in accuracy, proving that custom features tailored to cryptographic mechanics are more effective than generic statistical tests alone.
Conclusion and Future Work
While 58.5% accuracy is a significant leap forward, it also underscores the inherent difficulty of the task; many classical ciphers produce statistically similar outputs. The authors envision this tool as a "recommender" for human cryptanalysts, providing a ranked list of likely cipher types. Future work may involve expanding the model to support non-English languages and exploring deep learning architectures to further refine detection capabilities.
Summary updated