Deep Learning Fuzz Testing Methods for Unstructured Case
Haotian Yu, Xiaoguang Li, Yuefeng Du · 2022
Fuzz testing is to modify a large number of test cases by variation to find out the anomalies and vulnerabilities in target program. The Seq2Seq based fuzzing model, can automatically generates test cases that may cover the new path by learning the relationship between the initial seed cases and the corresponding execution path, and thus achieves better fuzz testing performance for structured cases. For unstructured cases, however the Seq2Seq model has two problems. In this paper, we propose a new deep learning testing model based on attention and scheduled sampling for unstructured cases. The model introduces an attention probability distribution on the input information sequences, and adopts decoders twice to generate input sequences by mixing the real information sequences with the predicted sequences. The model improves the coverage of execution paths in target program, and triggers more bugs.