Robust Asymmetric Heterogeneous Federated Learning With Corrupted Clients

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated learning (FL) systems face significant challenges from data corruption and client model heterogeneity.
• Data corruption, caused by noise or compression, severely degrades FL system performance.
• Existing FL methods struggle with diverse client models and unreliable data.

Purpose of the Study:

• To develop a novel framework for robust federated learning that addresses both model heterogeneity and data corruption.
• To enhance the resilience and adaptability of local models against various data corruption patterns.
• To mitigate the impact of corrupt feedback from less robust clients in collaborative learning.

Main Methods:

• Introduced a Robust Asymmetric Heterogeneous Federated Learning (RAHFL) framework.
• Proposed a Diversity-enhanced supervised Contrastive Learning technique using mixed-data augmentation for robust feature representation.
• Designed an Asymmetric Heterogeneous Federated Learning strategy enabling selective one-way learning to filter low-quality information.

Main Results:

• The proposed Diversity-enhanced supervised Contrastive Learning significantly improves model resilience to data corruption.
• The Asymmetric Heterogeneous Federated Learning strategy effectively prevents the propagation of corrupt information from underperforming clients.
• Extensive experiments validated the RAHFL framework's effectiveness and robustness in challenging, diverse federated learning scenarios.

Conclusions:

• The RAHFL framework offers a robust solution for federated learning with heterogeneous and corrupted client data.
• The combination of contrastive learning and asymmetric learning strategies enhances overall system reliability and performance.
• This approach represents a significant advancement in building dependable federated learning systems for real-world applications.