[S&P'19] NEUZZ: Efficient Fuzzing with Neural Program Smoothing

One of the main limitations of evolutionary optimization algorithms is that they cannot leverage the structure (i.e., gradients or other higher-order derivatives) of the underlying optimization problem. In this paper, we introduce a novel, efficient, and scalable program smoothing technique using feed-forward Neural Networks (NNs) that can incrementally learn smooth approximations of complex, real-world program branching behaviors, i.e., predicting the control flow edges of the target program exercised by a particular given input.

The key challenge in incremental training is that if an NN is only trained on new data, it might completely forget the rules it learned from old data [57]. We avoid this problem by designing a new coverage-based filtration scheme that creates a condensed summary of both old and new data, allowing the NN to be trained efficiently on them.