Saturday 08 March 2025
A novel approach to federated learning has been proposed, which aims to tackle the challenge of domain shifts in decentralized machine learning. Federated learning is a distributed AI technique that enables multiple clients to collaboratively train a shared model without sharing their individual data. However, this method often struggles when the clients’ data distributions differ significantly, a problem known as domain shift.
To address this issue, researchers have developed a new framework called I2PFL (Intra- and Inter-Domain Prototypes for Federated Learning). This approach combines two key components: intra-domain prototypes and inter-domain prototypes. The former are used to learn local features within each client’s data distribution, while the latter are employed to capture the relationships between clients’ data distributions.
The I2PFL framework begins by initializing a shared global model at the server, which is then distributed to each client. Each client trains their local model using their own data and computes intra-domain prototypes. These prototypes are augmented with MixUp, a technique that combines multiple examples from different classes, to enhance feature diversity. The clients also receive inter-domain prototypes from the server, which represent the shared knowledge across all clients.
During the training process, each client uses both intra- and inter-domain prototypes to update their local model. The intra-domain prototypes help the clients adapt to their local data distribution, while the inter-domain prototypes provide a broader perspective on how to generalize to other domains. This dual approach enables I2PFL to learn more robust and transferable features.
To evaluate the effectiveness of I2PFL, the researchers conducted extensive experiments on three image classification datasets: Digits, Office-10, and PACS. The results show that I2PFL significantly outperforms existing federated learning methods in terms of accuracy and robustness to domain shifts. For instance, on the Office-10 dataset, I2PFL achieved an average accuracy of 69.68%, while the best baseline method only managed 65.88%.
The authors also conducted a series of ablation studies to analyze the impact of different components within the I2PFL framework. The results suggest that both intra-domain prototypes and inter-domain prototypes are crucial for achieving good performance. Furthermore, the EMA (Exponential Moving Average) update mechanism used in I2PFL helps to stabilize the learning process and improve the overall robustness.
Cite this article: “Federated Learning with Domain Adaptation using Intra- and Inter-Domain Prototypes (I2PFL)”, The Science Archive, 2025.
Federated Learning, Domain Shift, Prototypes, Intra-Domain, Inter-Domain, Mixup, Image Classification, Digits, Office-10, Pacs
