There are still many challenges in Federated Learning (FL). First, during the model update process, the model parameters on the local user need to be sent to the server for aggregation. This involves the consumption of network bandwidth, especially when the number of users participating in FL is large. High communication costs may limit the application of FL in certain scenarios. Secondly, since users participating in FL usually have different data distributions, this heterogeneity of data may lead to poor model performance or even failure to converge. Third, privacy and security issues are also challenges that need to be addressed in FL. There is still a risk of information leakage during model aggregation. Malicious users may obtain sensitive information by analyzing communications during model updates or aggregation processes. To address these challenges, we propose HyperFedNet (HFN), an innovative approach that leverages hypernetwork. HFN introduces a paradigm shift in transmission aggregation within FL. Unlike traditional FL methods that transmit a large number of parameters from the main network, HFN reduces the communication burden and improves security by transmitting a compact set of hypernetwork parameters. After the parameters of the hypernetwork are deployed locally to the user, the local database features quantified by the embedding vector can be used as input, and parameters can be dynamically generated for the FL main network through user forward propagation. HFN efficiently reduces communication costs while improving accuracy. Extensive experimentation demonstrates that HFN outperforms traditional FL methods significantly. By seamlessly integrating this concept into the conventional FL algorithm, we achieve even more impressive results compared to the original approach.

翻译：联邦学习仍面临诸多挑战。首先，在模型更新过程中，本地用户的模型参数需发送至服务器进行聚合，这会消耗网络带宽，尤其在参与联邦学习的用户规模庞大时更为显著。高昂的通信成本可能限制联邦学习在特定场景中的应用。其次，由于参与联邦学习的用户通常具有不同的数据分布，这种数据异质性可能导致模型性能下降甚至无法收敛。第三，隐私与安全问题同样是联邦学习亟需解决的挑战——模型聚合过程中仍存在信息泄露风险，恶意用户可能通过分析模型更新或聚合过程中的通信量来获取敏感信息。针对这些挑战，我们提出HyperFedNet（HFN），一种创新性的超网络方法。HFN引入了联邦学习中传输聚合的范式转变：不同于传统联邦学习方法传输主网络的大量参数，HFN通过传输紧凑的超网络参数集降低了通信负担并提升了安全性。当超网络参数被本地化部署至用户端后，可通过嵌入向量量化的本地数据库特征作为输入，经由用户前向传播动态生成联邦学习主网络参数。HFN在提升准确率的同时有效降低了通信成本。大量实验表明，HFN显著优于传统联邦学习方法。通过将这一理念无缝集成至常规联邦学习算法，我们相较于原始方法取得了更为卓著的成果。