Towards Automatic Software Lineage Inference

Software continuously evolves to reflect changing requirements, feature updates, and bug fixes. Most existing research focuses on analyzing software release histories to understand the software evolution process and to describe evolutionary relationships among programs. However, there has been little research on inferring software lineage from (binary) programs. In this paper, we take a systematic approach towards software lineage inference. We explore three fundamental questions not addressed by existing work. First, how do we measure the quality of a lineage inference algorithm? Second, given existing approaches to binary similarity analysis, how good are they for lineage both currently and in an idealized setting? Third, what are the challenging problems in software lineage inference? Towards these goals we build LIMetric—a system for automatic software lineage inference of binary programs. We evaluated LIMetric on two types of lineage—straight line lineage and directed acyclic graph (DAG) lineage. We have also extended our technique to handle multiple straight line lineages. Our experiments used large scale real-world programs, with a total of 1,777 releases spanning over a combined 110 years of development history. In order to quantify lineage quality, we propose four metrics: (i) number of inversions and (ii) edit distance to monotonicity for straight line lineage, and (iii) number of lowest common ancestor (LCA) mismatches and (iv) average pairwise distance to true LCA for DAG lineage. LIMetric effectively extracted software derivation relationships among binary programs with high accuracy. Through close case analysis, we also formulate several challenging problems in software lineage inference that need to be addressed to attain even higher accuracy. Keywords-software evolution, software lineage, systematic evaluation