基于强化学习算法的神经网络模糊测试技术优化研究

Research of Neural Network Fuzz Testing Method Based on Reinforcement

摘要:

现有神经网络模糊测试技术在测试样本生成阶段通常对初始样本进行随机变异，导致生成样本质量不高，从而测试覆盖率不高。针对以上问题，提出一种基于强化学习算法的神经网络模糊测试技术，将模糊测试过程建模为马尔可夫决策过程，在该模型中，测试样本被看作环境状态，不同的变异方法被看作可供选择的动作空间，神经元覆盖率被看作奖励反馈，使用强化学习算法来学习最优的变异策略，指导生成最优测试样本，使其能够获得最高的神经元覆盖率。通过与现有的主流神经网络模糊测试方法的对比实验表明，基于强化学习算法的神经网络模糊测试技术，可以提升在不同粒度下的神经元覆盖。

Abstract:

The existing neural network fuzz testing techniques typically apply random mutations to initial samples during the test sample generation phase, resulting in low-quality generated samples and, consequently, low test coverage. To address these issues, we propose a neural network fuzz testing technique based on reinforcement learning algorithms. In this approach, we model the fuzz testing process as a Markov decision process. In this model, test samples are regarded as environmental states, and different mutation methods form a set of available actions, neuron coverage serves as a reward feedback Reinforcement learning algorithms are used to learn the optimal mutation strategy, guiding the generation of optimal test samples to achieve highest neuron coverage. Comparative experiments with mainstream neural network fuzz testing methods demonstrate that the neural network fuzz testing technique based on reinforcement learning algorithms can enhance neuron coverage at different granularities.