This paper analyzes the use of machine learning techniques for the identification of encryption algorithms, from ciphertexts only. The experiment involved corpora of plain texts in seven different languages; seven encryption algorithms, each one in ECB and CBC modes; and six data mining algorithms for classification. The plain text files were encrypted with each cryptographic algorithm under both cipher modes. After that, the ciphertexts were processed to produce metadata, which were then used by the classification algorithms. The overall experiment involved not only a high quantity of ciphertexts, but also time consuming procedures for metadata creation as well as for identification. Therefore, a high performance computer and customized memory management were employed. As expected, the results for ECB mode encryption algorithm identification were significantly high, and also reached full recognition. On the other hand, algorithm identification under CBC is supposed to be marginal, but successful identification was up to six times higher than the probabilistic bid.

cryptographic algorithm identification, data mining, machine intelligence, parallel computing