Evaluating Deep Learning Models to Generate Test Data In File Based Fuzzers

Fuzzing means repeatedly running program under test by modified inputs, with the aim of finding vulnerabilities in it. If program inputs under test have a complex structure, generating modified inputs for fuzzing is not an easy task. Best solution in these cases is to use the input structure of the program under test to produce accurate test data. The problem is that maybe the input structure documentation of program under test not be available. Human understanding of such complex structures is also very difficult, costly, time consuming, and prone to human errors. To overcome to above problems, use of machine learning and deep neural networks to automatically learn the complex structures of program inputs and generate test data tailored to this structure has been proposed. One of main challenges in this field is using appropriate deep learning model to intended usage. In this paper, suitable deep learning models for learning and test data generation in file-based fuzzers are studied. Also by introducing appropriate parameters to performance evaluation, the evaluation has performed. So recurrent neural networks and its derivations introduced as best deep learning models for text data. Also, effective parameters to performance evalution include training time, loss value in training and evaluate time. Loss value as main parameter once used in various deep learning models with same structure and again in same deep learning models with various structures to select best deep learning model.