Related Experiment Videos
A unified basis decomposition framework for addressing the federated learning trilemma: Communication-efficiency,
Zhaoyang Ma1, Zhihao Wu2, Juncheng Wang3
1Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science and Technology, Beijing Jiaotong University, Beijing,100044, China; School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore, 639798, Singapore.
Summary
Federated Basis Decomposition (FedBD) offers a unified solution to the federated learning trilemma. This novel framework balances communication efficiency, personalization, and privacy, outperforming traditional methods in medical imaging tasks.
Area of Science:
- Machine Learning
- Artificial Intelligence
- Medical Imaging
Background:
- Federated learning (FL) faces a critical challenge in balancing communication efficiency, personalized adaptation, and privacy protection.
- Existing FL methods often compromise one aspect to optimize others, lacking a holistic approach.
- This limitation hinders the practical application of FL in sensitive domains like healthcare.
Purpose of the Study:
- To introduce Federated Basis Decomposition (FedBD), a unified framework designed to address the FL trilemma.
- To achieve simultaneous optimization of communication efficiency, personalization, and privacy.
- To enable practical FL deployment in complex, real-world scenarios, particularly in medical imaging.
Main Methods:
- FedBD reformulates federated learning as a subspace optimization problem using layer-wise basis decomposition.
- Clients exchange low-dimensional scalar weights derived from globally shared basis models, reducing communication overhead.
- Privacy is enhanced by projecting gradients into a random basis subspace, and personalization is achieved by decoupling shared weights from local parameters.
Main Results:
- FedBD demonstrated a significant accuracy density gain of up to 9.73× on three medical imaging datasets.
- The framework achieved an order-of-magnitude reduction in communication overhead compared to conventional FL methods.
- Competitive accuracy was maintained while drastically improving communication efficiency and privacy.
Conclusions:
- FedBD provides a unified solution to the fundamental trilemma in federated learning.
- The proposed method effectively balances communication efficiency, personalization, and privacy.
- FedBD shows significant promise for practical FL applications in resource-constrained and privacy-sensitive environments.
