Abstract: The Internet of things (IoT) has shown a wide range of application prospects and huge development potential in many fields. However, as the scale of the IoT continues to expand, independent IoT devices lack highquality attack instances, making it difficult to effectively respond to increasingly complex and diverse attack behaviors. Consequently, addressing IoT security issues has become a critical challenge that requires urgent attention. To address this problem, the paper proposes an IoT intrusion detection model based on federated learning and attention mechanisms, which allows multiple devices to train the global model collaboratively while protecting their data privacy. Firstly, this paper constructs an intrusion detection model combining convolutional neural network and mixed attention mechanism to extract key features of network traffic data, so as to improve detection accuracy. Secondly, the paper introduces the model contrast loss to correct the training direction of the local model to alleviate the global model convergence difficulties caused by the nonindependent and same distribution of data between devices. The experimental results show that the proposed model is significantly superior to the existing methods in terms of accuracy, accuracy and recall, demonstrating stronger intrusion detection capabilities, and can effectively deal with complex data distribution problems in the IoT environment.

Key words: federated learning, Internet of things security, intrusion detection, deep learning, attention mechanisms

摘要： 物联网在众多领域中展现出广泛的应用前景和巨大的发展潜力.然而，随着物联网规模的持续扩展，独立的物联网设备缺乏高质量攻击实例，难以有效应对日益复杂且多样化的攻击行为，物联网安全问题已经成为亟待解决的关键挑战.为应对这一问题，提出了一种基于联邦学习和注意力机制的物联网入侵检测模型，允许多个设备在保护其数据隐私的基础上协同训练全局模型.首先，构建了一个结合卷积神经网络与混合注意力机制的入侵检测模型，提取网络流量数据的关键特征，从而提高检测的准确率.其次，引入模型对比损失，通过矫正本地模型的训练方向，缓解设备间数据非独立同分布所导致的全局模型收敛困难等问题.实验结果显示，该模型在准确率、精确率和召回率等指标上显著优于现有方法，展现了更强的入侵检测能力，能够有效应对物联网环境中复杂的数据分布问题.

关键词: 联邦学习, 物联网安全, 入侵检测, 深度学习, 注意力机制