Layer-based personalized multi-fusion federated learning

Historical data form the foundation for information system modelling. Yet, clients typically do not share it due to concerns about data privacy and security.In the absence of data, it is impossible to identify a model that yields optimal decisions for all clients. To overcome these challenges, this paper proposes a multi-layer multi-fusion strategy federated learning framework, where the server deploys different fusion strategies for neural network layers of client models. The corresponding fusion strategy for each neural network layer is selected based on its functions. Functionally, we categorised the neural network into a generalised feature extraction layer and a personalised fully-connected layer. Then, using an exponential similarity metric, the fusion weights for the fully connected layer corresponding to each client are calculated, thereby enhancing the efficiency of collaborative personalisation across heterogeneous data. Meanwhile, the feature extraction layer adopts the federated global optimal model approximation fusion strategy. Finally, extensive experimental results demonstrate that the proposed method outperforms existing comparable approaches, such as improving the recognition accuracy of CIFAR-100 from 0.4792 to 0.4859.