脑机接口作为人机交互的重要技术路径，在医疗康复、无障碍通信、神经调控等多个领域展现出广阔的应用前景。然而，脑机接口系统高度依赖机器学习对脑电信号进行解码，近年来暴露的对抗攻击问题使得其安全性面临严峻挑战。对抗攻击通过在输入信号中添加微小扰动，即可诱导模型产生错误预测，进而可能导致设备失控或被远程操控，严重威胁系统可靠性与用户人身安全。本文聚焦非侵入式脑机接口中的对抗安全问题，系统梳理当前研究进展，从攻击与防御两个维度展开综述。通过全面评估现有成果，本文总结了当前脑机接口系统在安全性方面存在的关键挑战，并提出未来值得关注的研究方向，为构建安全、可靠、可部署的脑机接口系统提供理论支持与技术参考。

 Brain-computer interfaces (BCIs) have shown great potential in rehabilitation, assistive communication, and neuromodulation. However, their strong reliance on machine learning for EEG decoding exposes them to adversarial attacks. These attacks add subtle perturbations to input signals, which can mislead models and cause incorrect predictions, leading to system malfunction or even remote manipulation. This paper focuses on adversarial security in non-invasive BCIs, providing a systematic review of current research on both attacks and defences. It summarises key challenges and outlines future directions to support the development of secure and reliable BCI systems.