基于语义相似性的联邦对比学习

Federated Contrastive Learning Based on Semantic Similarity

摘要: 联邦学习在数据异构场景中面临着数据分布偏斜以及知识积累效率低下的双重挑战，尽管研究者们已提出诸如原型对比学习和知识蒸馏等方法来应对这些挑战，但这些方法未能充分考虑语义相似性的影响，从而导致语义相似的类别难以区分、全局原型质量欠佳以及知识积累效率低下。为解决这些问题，文章提出了一种基于语义相似性的联邦对比学习框架。该框架通过结合原型相似性和数据量聚合全局原型，为后续训练提供高质量的基础。然后利用全局语义关联矩阵指导知识蒸馏，高效地积累共性知识。最后，使用全局语义关联矩阵动态筛选困难负原型进行对比学习，以精细化决策边界。实验结果表明，与现有算法相比，本文提出的算法在准确率上提升了6.7%至7.7%，罕见类别的召回率和F1值提升了20%，提高了系统在数据异构场景下的泛化性能。

Abstract: Federated learning faces the dual challenges of data distribution skew and low knowledge accumulation efficiency in heterogeneous data scenarios. Although researchers have proposed methods such as prototype contrastive learning and knowledge distillation to address these challenges, these methods fail to fully consider the impact of semantic similarity. As a result, semantically similar categories are difficult to distinguish, the quality of global prototypes is suboptimal, and knowledge accumulation efficiency remains low. To address these issues, this paper proposes a federated contrastive learning framework based on semantic similarity. The framework combines prototype similarity and data volume to aggregate global prototypes, providing a high-quality foundation for subsequent training. It then utilizes the global semantic association matrix to guide knowledge distillation, efficiently accumulating common knowledge. Finally, it dynamically screens difficult negative prototypes for contrastive learning using the global semantic association matrix to refine decision boundaries. Experimental results show that compared with existing algorithms, the proposed algorithm in this paper improves accuracy by 6.7% to 7.7%, and the recall and F1 values of rare categories are increased by 20%, thereby enhancing the system’s generalization performance in heterogeneous data scenarios.