MLPNeuzz: A Novel Neural Program Smoothing Method Based on Multi-Layer Perceptron

Abstract

In recent years, using fuzzy methods to mine network security vulnerabilities has become a mainstream. Fuzzing is an effective vulnerability mining technology, which can find the potential vulnerability trigger point by traversing the program branch through some key algorithms. However, the traditional fuzzing methods exist some problems, such as redundant test cases, inefficient mutation strategy and so on. Therefore, a method combining machine learning with fuzzing has been proposed, which provides solutions to the above problems. Recently, someone proposes an effective fuzzer called NEUZZ, which uses a simple feedforward neural network (FNN) for neural program smoothing to model the branching behavior of the target program and improve the utilization of test cases. However, the traditional FNN model is easy to cause low learning efficiency and poor generalization ability and other problems. In order to solve these problems, a novel neural program smoothing method based on Multi-Layer Perceptron (MLP) is proposed in this paper, and we name the fuzzer as MLPNeuzz. MLPNeuzz can further collect edge coverage information and improve the smoothing effect of neural programs. In addition, we refine the original NEUZZ fuzzy method to make its fuzzy process more reasonable. Experiments on several real-world application programs show that the MLPNeuzz method proposed in this paper can achieve higher edge coverage than NEUZZ under the same time overhead.