To address the issues of low initial seed efficiency and a large number of ineffective mutations, this paper proposes an innovative fuzz testing seed optimization method combining neural networks and genetic algorithms. Traditional fuzz testing seed generation typically relies on random selection and the number of covered paths. In contrast, our method significantly improves seed generation efficiency and coverage by incorporating neural network models and genetic algorithms. First, the AFL tool is used to generate seed coverage path data, which is then used to train the neural network model. This model is employed to construct a fitness function to assess the potential of each seed. Subsequently, new seeds are generated through genetic algorithm crossover and mutation operations, with fitness evaluations based on the predictions of the neural network. Ultimately, the genetic algorithm optimizes the seeds through multiple generations, progressively improving coverage and vulnerability discovery capabilities. The experimental results demonstrate that the proposed method achieves significant improvements in fuzz testing performance, with path coverage increased by 28% compared to AFL and 23% compared to AFL++, and vulnerability discovery enhanced by over 200%.
Jiang et al. (Fri,) studied this question.