【摘要】 为了解决物联网设备在资源受限和数据非独立同分布（Non-IID）时出现的训练效率低、模型性能差的问题，提出了一种个性化剪枝联邦学习框架用于物联网的入侵检测。首先，提出了一种基于通道重要性评分的结构化剪枝策略，该策略通过平衡模型的准确率与复杂度来生成子模型下发给资源受限客户端。其次，提出了一种异构模型聚合算法，对通道采用相似度加权系数进行加权平均，有效降低了Non-IID数据在模型聚合中的负面影响。最后，网络入侵数据集BoT-IoT上的实验结果表明，相较于现有方法，所提方法能显著降低资源受限客户端的时间开销，处理速度提升20.82%，并且在Non-IID场景下，入侵检测的准确率提高0.86%。更多还原

【Abstract】 In order to address the issue of inadequate training efficiency and subpar model performance encountered by Internet of things(Io T) devices when dealing with resource constraints and non-independent and identically distributed(Non-IID) data, a novel personalized pruning federated learning frame work for IoT intrusion detection was put forth. Initially, a channel importance scoring-based structured pruning strategy was proposed, facilitating the generation of submodels to be disseminated to resource-limited clients, thereby harmonizing model accuracy and complexity. Subsequently,an innovative heterogeneous model aggregation algorithm was introduced, utilizing similarity-weighted coefficients for channel averaging, thereby effectively mitigating the adverse effects of Non-IID data during the model aggregation process. Ultimately, experimental results derived from the network intrusion dataset BoT-IoT substantiate that, relative to existing methods, the proposed method notably curtails the time expenditure of resource-constrained clients, and improves processing speed by 20.82%, while enhancing the accuracy of intrusion detection by 0.86% in Non-IID conditions.更多还原