The ICT-CAS paper MetaFed: Federated Learning among Federations with Cyclic Knowledge Distillation for Personalized Healthcare, from ICT’s Research Center for Ubiquitous Computing Systems, won the Innovation Award at the workshop, FL-IJCAI'22. This workshop provides an important platform for researchers, practitioners, and industry experts to share and discuss the latest advances and trends in the field of federated learning.
As outlined in the paper, MetaFed is a novel meta federated learning framework for personalized federated learning among federations. Federated learning has attracted increasing attention to building models without accessing the raw user data, especially important in healthcare. In real applications, clients may form multiple federations which have decentralized requirements for federated learning. If distribution heterogeneity between federations makes it difficult to directly apply existing methods, the challenge lies in effectively integrating federated knowledge and obtaining personalized models.
To address such issues, ICT’s Research Center for Ubiquitous Computing Systems proposed MetaFed, a novel meta federated learning framework via cyclic knowledge distillation for healthcare. It can accumulate common information from different federations without compromising privacy security, and then achieve personalized models for each federation through adapted knowledge distillation. Specifically, MetaFed treats each federation as a meta distribution and aggregates knowledge of each federation in a cyclic manner. Training is in two parts: common knowledge accumulation and personalization. Comprehensive experiments on image and time-series datasets illustrate that MetaFed has remarkable performance in each federation without a server, as compared to state-of-the-art methods (e.g., 10%+ accuracy improvement compared to the baseline for PAMAP2). MetaFed reduces the number of rounds, thus saving communication costs. It is extensible, can be used in many healthcare applications, and can work well in many circumstances.
MetaFed treats each federation as a meta distribution
The Research Center for Ubiquitous Computing Systems has focused for many years on developing technologies and systems for ubiquitous computing, especially in healthcare-related fields. The center proposed the first federated transfer learning framework, FedHealth, for wearable healthcare, and it designed PdAssist, a well-performed objective and quantified symptom assessment tool of PD on mobile devices. In the era when data privacy and security are increasingly emphasized, MetaFed has proposed a novel federated scenario and an excellent yet simple solution, making it possible to apply federated learning in practical healthcare settings.
The FL-IJCAI'22, focusing on knowledge-sharing and collaboration in the field of federated learning, covers a broad range of topics, including privacy and security issues, optimization techniques, and real-world applications of trustworthy federated learning.