A Survey of Security Aggregation
A Survey of Security Aggregation
复制标题
DOI：
10.23919/icact53585.2022.9728912
复制
发表时间：
2022-02
期刊：
影响因子：
--
通讯作者：
Sufang Zhou;Mingzhe Liao;Baojun Qiao;Xiaobo Yang
中科院分区：
文献类型：
--
作者：
Sufang Zhou;Mingzhe Liao;Baojun Qiao;Xiaobo Yang
机器学习（ML）需要收集大量数据来训练强大的预测模型。联邦学习（FL）允许将敏感数据保存在客户端以训练共享模型，但共享模型仍然会带来隐私问题。安全聚合是一种重要的算法，它能够计算客户端模型更新的总和，而不会泄露有关单个客户端更新的信息，从而确保联合学习的安全。在这项研究中，我们调查了近年来的安全聚集协议的结果，并根据秘密共享，差分隐私和同态加密机制的安全聚集领域的研究成果进行了综述。在此基础上，比较分析了不同机制的优缺点，并从安全性、动态用户鲁棒性、计算代价和通信代价等方面对安全聚合协议进行了评价。最后，我们对安全聚合协议的未来发展进行了展望，并给出了未来可能的研究方向。
Machine learning (ML) requires the collection of large amounts of data to train robust predictive models. Federated learning (FL) allows sensitive data to be kept on the client side to train a shared model, but shared models still pose privacy concerns. Secure aggregation is an important algorithm for securing federation learning by being able to compute the sum of client-side model updates without revealing information about individual client updates. In this study, we investigate the results of secure aggregation protocols in recent years and review the research results in the field of secure aggregation according to secret sharing, differential privacy, and homomorphic encryption mechanisms. On this basis, we compare and analyze the advantages and disadvantages of different mechanisms and then evaluate the security aggregation protocols in terms of security, dynamic user robustness, computational cost, and communication cost. Finally, we provide an outlook on the future development of secure aggregation protocols and give possible future research directions.