We present a novel privacy-preserving model aggregation for asynchronous federated learning, named PPA-AFL that removes the restriction of synchronous aggregation of local model updates in federated learning, while enabling the protection of the local model updates against the server. In PPA-AFL, clients can proactive decide when to engage in the training process, and sends local model updates to the server when the updates are available. Thus, it is not necessary to keep synchronicity with other clients. To safeguard client updates and facilitate local model aggregation, we employ Paillier encryption for local update encryption and support homomorphic aggregation. Furthermore, secret sharing is utilized to enable the sharing of decryption keys and facilitate privacy-preserving asynchronous aggregation. As a result, the server remains unable to gain any information about the local updates while asynchronously aggregating to produce the global model. We demonstrate the efficacy of our proposed PPA-AFL framework through comprehensive complexity analysis and extensive experiments on a prototype implementation, highlighting its potential for practical adoption in privacy-sensitive asynchronous federated learning scenarios.

翻译：本文提出了一种面向异步联邦学习的隐私保护模型聚合框架，命名为PPA-AFL。该框架消除了联邦学习中本地模型更新需同步聚合的限制，同时实现对服务器端保护的本地模型更新。在PPA-AFL中，客户端可以主动决定何时参与训练过程，并在本地模型更新可用时将其发送至服务器，因此无需与其他客户端保持同步。为保护客户端更新并支持本地模型聚合，我们采用Paillier加密机制对本地更新进行加密，并支持同态聚合操作。此外，通过引入秘密共享技术实现解密密钥的共享，从而支持隐私保护的异步聚合。最终，服务器在异步聚合生成全局模型的过程中，无法获取关于本地更新的任何信息。通过全面的复杂度分析及原型系统上的大量实验，我们验证了所提PPA-AFL框架的有效性，凸显了其在隐私敏感的异步联邦学习场景中实际应用的潜力。