Automatically generating models for data flow analysis using machine learning

A method comprising:

• acquiring a set of function stubs of a software program;
• acquiring source code for a set of functions corresponding to the set of function stubs and library documentation related to the set of functions;
• executing a query model that generates a set of queries, wherein the query model bases the set of queries on the source code and the library documentation so that the set of queries is tailored to a context of the software program;
• executing a machine learning model that takes the source code, the library documentation, and the set of queries as inputs and generates as an output a set of answers to the set of queries;
• constructing a function model of a function of the set of functions, using the set of answers;
• and
• executing a data flow analysis of the software program, wherein the data flow analysis uses the function model to trace a flow of test data through the source code, and wherein the function model describes a purpose of the function within the context of the software program in a manner that is understandable by a data flow engine that performs the data flow analysis.

A system comprising:

• a processor;
• and
• a non-transitory computer readable medium comprising stored instructions which, when executed by the processor, cause the processor to:
• acquire a first function model of a source code element, wherein the first function model was automatically generated using answers output by a machine learning technique in response to a set of inputs comprising:
• source code for a set of functions from which the source code element was taken, library documentation related to the set of functions, and a set of queries tailored to a context of the source code element based on the source code and the library documentation;
• acquire a second function model of the source code element, wherein the second function model was manually generated by a human;
• convert the second function model into a set of labels;
• compare the set of labels to another set of labels from which the first function model was generated, to determine a degree of match;
• and
• initiate a retraining of a machine learning model used for the machine learning technique when the degree of match is less than a threshold, wherein the retraining utilizes new training data to retrain the machine learning model so that the degree of match at least meets the threshold.