Neural Fuzzing: A Neural Approach to Generate Test Data for File Format Fuzzing
share
This article is aimed at the design and implementation of a file format fuzzer. Files are significant inputs to the most of real-world applications. A substantial difficulty with generating input files as test data is to recon the underlying structure and format of the files. In order to distinguish pure data stored in a file from the meta-data describing the file format, a deep learning method based on a neural language model is proposed in this article. The resultant learned model could be applied as a hybrid test data generator, to generate and fuzz both the textual and none-textual sections of the input file. Moreover, the model could be applied to generate test data to fuzz both the meta-data and the ordinary data stored in the file. Our experiments with two known fuzzing tools, AFL and Learn&Fuzz, demonstrate the relatively high code coverage of our proposed method. The experiments also indicate simple neural language models provide a more accurate learning model, than the complicated encoder-decoder models.
READ FULL TEXT
