Model Weights Protection in Federated Learning using Hypernetworks
Shuting Li, Hong Liang Han, Yi Luo, Aiguo Chen · 2025
Federated Learning (FL) is an emerging distributed learning paradigm that facilitates local model training while exchanging only model parameter updates, thereby preserving individual data privacy.However, in typical FL settings, the sharing of global model knowledge with clients during the training process can lead to privacy leakage of the global model, potentially infringing upon the interests of the model owner.To address this challenge, this paper proposes a method, termed Generating Global model For Federated Learning(FedGG), that leverages hypernetworks to generate highquality global models with protected weights in FL.Specifically, We generate a privacy-protected global model at the server using a hypernetwork, which takes in a privacy-preserving data descriptor and outputs a set of weights tailored to the described data.The embedding network, responsible for generating the data descriptors, and the hypernetwork, which produces the model weights, are co-trained by clients using their local data in a federated learning framework.Upon training completion, the embedding network produces a global descriptor that captures the characteristics of all clients' data, while ensuring privacy through the use of datafree knowledge distillation.This global descriptor is subsequently fed into the hypernetwork to generate the global model's weights, which are kept private from all clients.By leveraging this approach, we ensure that the model owner's interests are effectively protected, maintaining data privacy and safeguarding proprietary information throughout the process.Experimental results on standard benchmarks demonstrate that, under the premise of model weight protection, the global model trained using FedGG outperforms baseline method.