【目的】针对新型电力系统数据驱动算法对抗攻击风险及攻防协同性不足的问题，搭建对抗攻击与防御协同优化理论框架，提升攻击靶向性、防御鲁棒性及复杂攻击特征辨识能力，建立攻防协同进化的闭环优化机制。【方法】在对抗攻击生成模块中，通过自注意力机制量化节点特征贡献度并结合Top-K策略筛选关键节点；利用编解码器与强化学习动态优化扰动策略，经过滤器保留关键节点扰动以提升攻击效率。在对抗防御模型中，采用栈式自编码器提取静态结构特征，卷积神经网络-长短期记忆网络融合时空特征，通过动态权重策略整合多模态特征后，经支持向量机分类器实现攻击样本与正常样本的辨识。【结果】相较于随机节点攻击、快速梯度符号法及投影梯度下降攻击方法，所提攻击方法在维持较高成功率的同时，其全攻击强度范围内的鲁棒性更贴合电力系统对抗攻击实际需求，且扰动可集中于关键节点，由此验证了攻击靶向性优势；防御层面，融合模型性能显著优于单一模型，凸显多模态特征融合对复杂攻击模式的强辨识能力。【结论】攻击侧融合自注意力与强化学习，实现了关键节点的靶向扰动；防御侧采用多模态特征融合，提升了复杂攻击的辨识能力；并通过闭环反馈机制，实现了攻防策略的动态协同进化。

[Objective] Aiming at the problems of adversarial attack risks and insufficient offensive-defense coordination of data-driven algorithms in new power systems， a theoretical framework for co-optimization of adversarial attacks and defense is established. This framework aims to enhance attack targeting， defense robustness， and the capability to identify complex attack features， thereby establishing a closed-loop optimization mechanism for offensive-defense co-evolution. [Methods] In the adversarial attack generation module， a self-attention mechanism is utilized to quantify node feature contributions， and a Top-K strategy is combined to screen key nodes. An encoder-decoder architecture and reinforcement learning are employed to dynamically optimize perturbation strategies， and a filter retains perturbations on key nodes to improve attack efficiency. In the adversarial defense model， a stacked autoencoder extracts static structural features， while a convolutional neural network-long short-term memory network fuses spatiotemporal features. These multi-modal features are then integrated via a dynamic weighting strategy and fed into a support vector machine classifier to distinguish attack samples from normal samples. [Results] Compared with random node attacks， the fast gradient sign method， and projected gradient descent attacks， the proposed attack method maintains a high success rate while demonstrating robustness across the entire attack intensity range that better aligns with the practical requirements of power system adversarial attacks. Furthermore， perturbations can be concentrated on key nodes， verifying the advantage of attack targeting. On the defense side， the fusion model's performance significantly surpasses that of single models， highlighting the strong identification capability of multi-modal feature fusion for complex attack patterns. [Conclusions] On the attack side， the integration of self-attention and reinforcement learning achieves targeted perturbation on key nodes. On the defense side， the adoption of multi-modal feature fusion enhances the identification capability for complex attacks. Furthermore， a dynamic co-evolution of offensive and defensive strategies is realized through a closed-loop feedback mechanism.