Personalized Fair Split Learning for Resource-Constrained Internet of Things.

Area of Science:

• Computer Science
• Artificial Intelligence
• Distributed Systems

Background:

• Federated learning (FL) is crucial for privacy-preserving data analysis in the Internet of Things (IoT).
• Resource-constrained IoT devices struggle with traditional FL due to limited computational power and storage.
• Data heterogeneity and unequal benefit distribution are significant challenges in IoT federated learning.

Purpose of the Study:

• To propose a personalized and fair split learning framework for resource-constrained IoT clients.
• To enable efficient model training and personalized adaptation on edge devices.
• To ensure equitable distribution of benefits among participating IoT devices.

Main Methods:

• A U-shaped model structure is employed, allowing clients to offload parts of the foundational model to a central server.
• Clients retain personalized model subsets locally for tailored requirements.
• An optimized model-aggregation method with contribution-based weights is used for fair benefit distribution.

Main Results:

• The proposed framework achieves higher accuracy compared to baseline methods across three data heterogeneity scenarios.
• The framework effectively addresses the challenges of limited resources and data heterogeneity in IoT environments.
• Collaborative fairness is successfully achieved, promoting balanced cooperation among devices.

Conclusions:

• The personalized and fair split learning framework offers a viable solution for federated learning on resource-constrained IoT devices.
• This approach enhances model performance and promotes equitable collaboration in decentralized AI.
• It paves the way for more sustainable and effective distributed learning in IoT ecosystems.