The Stacked Seq2seq-attention Model for Protocol Fuzzing

Fuzzing is an effective approach to discover vulnerabilities in software by generating the amount of unexcepted data as inputs to a program. It is difficult to fuzz the protocol automatically because it is necessary to manually construct a template that satisfies the protocol specification to generate test cases. In this paper, we establish stacked seq2seq-attention models to generate protocol test cases automatically. Seq2seq-attention is a machine learning technique which has an encoder-decoder structure to output text sequences based on context. We evaluate the training effect of seq2seq-attention models with different layers of LSTM and point out that the highest correctness of test cases is achieved by 3 layers LSTM. Besides, we implement a fuzzer based on stacked seq2seq attention model and compare with grammar-based fuzzer, which result indicates the test cases generated by our fuzzer discover more unique basic blocks.