Parametric optimization of input data array construction based on character cross-links

Abstract

The problem of parametric optimization of the input data array (IDA) construction process for source code authorship attribution is considered, in particular for distinguishing between human-written and large-language-model-generated code. A language-independent feature type has been further developed, namely character-level cross-links between words, represented as ordered pairs of word prefixes and suffixes within a sliding window. By considering all pairs of words, rather than only adjacent ones, these features are insensitive to local rearrangements of code fragments. The constructed input data array is evaluated using a tournament of 39 machine learning model architectures, and the classification quality is assessed both at the level of individual windows and at the author level by majority voting. A full-factorial computational experiment of 168 trials has been carried out for 12 authors (10 humans and 2 generative models) in four programming languages. The effects of window size, the informational sufficiency threshold, the filtering mode, and the cross-link rank on the proportion of correctly classified objects and the size of the feature dictionary have been studied. It has been established that cross-links improve performance in 37 % of the evaluated configurations. The greatest improvement (up to +7.56 %) is observed for a window size of 2,000 characters, which represents the point of maximum contribution of the additional features rather than the point of highest overall performance. The Pareto-optimal configuration uses a 500-character window, a threshold of three and the averaging filtering mode, achieving 99,37 % of correctly classified objects with a dictionary of only 310 features, whereas the maximum mode increases quality at the cost of a nearly twentyfold growth of the dictionary. The limits of applicability of the method and the conditions under which cross-links reduce classification quality are determined, in particular for authors with a high baseline due to the ceiling effect. Таbl.: 5. Figs.: 4. Refs.: 9 titles.