The analysis of dimensionality reduction techniques in cryptographic object code classification

This paper compares the application of three different dimension reduction techniques to the problem of classifying functions in object code form as being cryptographic in nature or not. A simple classifier is used to compare dimensionality reduction via sorted covariance, principal component analysis, and correlation-based feature subset selection. The analysis concentrates on the classification accuracy as the number of dimensions is increased. It is demonstrated that when discarding 90% of the measured dimensions, accuracy only suffers by 1% for this problem. By discarding dimensions, computational intelligence techniques can be applied with a drastic reduction in algorithmic complexity. The primary focus is on Intel IA32 instruction set, but analysis shows consistent results on the Sun SPARC instruction set.