Secure Hierarchical Asynchronous Federated Learning with Shuffle Model and Mask-DP.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Hierarchical asynchronous federated learning (HAFL) offers practical communication and aggregation but struggles with privacy and robustness.
• Malicious nodes in HAFL can compromise data privacy and poison global models, degrading system integrity.

Purpose of the Study:

• To propose a Secure Hierarchical Asynchronous Federated Learning (SHAFL) framework to address privacy and robustness challenges in HAFL.
• To enhance the security and performance of federated learning systems against malicious attacks and data inference.

Main Methods:

• SHAFL organizes nodes into groups with gateways, using mask-Differential Privacy (DP) exchange and homomorphic encryption (HE) to prevent inference and collusion attacks.
• SHAFL employs eliminable noise, a shuffle model, and subsampling to improve local model privacy and reduce noise impact on global performance.
• Global model aggregation in SHAFL balances accuracy and communication delay, mitigating the effects of malicious and stale models.

Main Results:

• SHAFL demonstrates superior convergence, security, robustness, and privacy-preserving capabilities compared to existing solutions.
• The framework effectively prevents inference attacks from gateways and committee nodes and collusion attacks among training nodes.
• SHAFL reduces the performance impact of noise and mitigates risks associated with malicious or delayed model updates.

Conclusions:

• The proposed SHAFL framework significantly enhances the security, robustness, and privacy of hierarchical asynchronous federated learning.
• SHAFL offers a practical solution for real-world federated learning applications requiring strong data protection and system integrity.
• SHAFL represents a state-of-the-art advancement in secure and robust federated learning systems.