Abstract: A federated learning scheme that balances efficiency and security has been designed to address the issue that implementing privacy protection mechanisms in federated learning will increase the communication burden on the system, and sacrificing model accuracy when attempting to improve system communication efficiency. This scheme aims to balance the communication efficiency and privacy security of the model, which is composed of time complexity and communication overhead. Firstly, the parameter sparsification strategy of the sparse ternary compression (STC) algorithm is used to compress the model parameters to be uploaded into triples to reduce the amount of data transmission; then, in order to compensate for the information loss caused by parameter compression, the error feedback mechanism is used to transmit the error as feedback information back to the model training; finally, the Paillier homomorphic encryption technology is used to ensure the privacy security of parameter transmission and aggregation under the premise of efficient model communication. The proposed scheme is compared with the current cutting-edge algorithms in Independent and Identically Distributed(IID) and Non Independent and Identically Distributed(Non-IID) data scenarios on multiple public datasets. The experimental results show that the proposed scheme solves the existing problem of being unable to strike a balance between time complexity, communication overhead and privacy protection, and has extremely high accuracy, convergence and robustness under three mainstream datasets and data distribution conditions.

Key words: federated learning, Sparse Ternary Compression(STC), communication efficiency, homomorphic encryption, privacy protection

摘要： 针对在联邦学习中实施隐私保护机制会加剧系统通信负担，而当试图提升系统通信效率时，又会牺牲模型精度的问题，设计了一种兼顾高效性和安全性的联邦学习方案。该方案旨在平衡由时间复杂度、通信开销构成的模型通信效率和隐私安全性。首先利用稀疏三元压缩(STC)算法的参数稀疏化策略将待上传的模型参数压缩为三元组形式，以减少数据传输量；接着，为弥补因参数压缩带来的信息损失，用错误反馈机制将误差作为反馈信息传回模型训练；最后，采用Paillier同态加密技术保证了模型高效通信前提下参数传输及聚合过程中的隐私安全。在多个公开数据集上将所提方案与当前前沿算法在独立同分布(IID)及非独立同分布(Non-IID)的数据场景下进行了对比分析，实验结果表明，所提方案解决了现存无法在时间复杂度、通信开销、隐私保护间取得平衡的问题，且在三个主流数据集和数据分布条件下都具有极高精度、收敛性及鲁棒性。

关键词: 联邦学习, 稀疏三元压缩, 通信效率, 同态加密, 隐私保护