Using classifiers to predict linear feedback shift registers

Previously (J.C. Hernandez et al., 2000), some new ideas that justify the use of artificial intelligence techniques in cryptanalysis are presented. The main objective of that paper was to show that the theoretical next bit prediction problem can be transformed into a classification problem, and this classification problem could be solved with the aid of some AI algorithms. In particular, they showed how a well-known classifier called c4.5 could predict the next bit generated by a linear feedback shift register (LFSR, a widely used model of pseudorandom number generator) very efficiently and, most importantly, without any previous knowledge over the model used. The authors look for other classifiers, apart from c4.5, that could be useful in the prediction of LFSRs. We conclude that the selection of c4.5 by Hernandez et al. was adequate, because it shows the best accuracy of all the classifiers tested. However, we have found other classifiers that produce interesting results, and we suggest that these algorithms must be taken into account in the future when trying to predict more complex LFSR-based models. Finally, we show some other properties that make the c4.5 algorithm the best choice for this particular cryptanalytic problem