An adversarial example defense method and device based on a model inversion method are disclosed. In order to solve the problem of lacking a low-cost and efficient adversarial example defense method in the field of deep neural network security, a model inversion mechanism based on a StyleGAN generator is proposed to realize an adversarial example defense method. Through the in-depth analysis of the generator StyleGAN, this paper proposes enhanced information training and improved proAdaIN, which is innovativly applied to the feature generation scheme of the adversarial example defense system. By adding noise, feature decoupling, and using conflict semantics to distinguish real samples from adversarial samples, this paper proposes a novel feature generation scheme for the adversarial example defense system. It solves the problems of high cost, low efficiency and poor defense effect in traditional defense schemes. 本发明公开的基于模型反演方法的对抗样本防御方法和设备，为了解决深度神经网络安全领域中缺少低成本、高效的对抗样本防御方法的问题，提出了一种基于StyleGAN生成器的模型反演机制实现对抗样本防御方法。通过对生成器StyleGAN的深入分析，提出强化信息训练和改进的proAdaIN，将其创新性地应用到对抗样本防御系统的特征生成方案中，并通过添加噪声、特征解耦、利用冲突语义区分真实样本和对抗样本，解决了传统防御方案中成本高、效率低、防御效果差等问题。