Quantization-based chained privacy-preserving federated learning.

Area of Science:

• Distributed Machine Learning
• Data Privacy and Security

Background:

• Federated Learning (FL) protects data privacy by training models locally.
• Traditional FL faces challenges with communication efficiency, computational costs, and privacy preservation, especially in edge computing.
• High overhead hinders real-time applications in current FL schemes.

Purpose of the Study:

• To propose an innovative federated learning framework, Q-Chain FL.
• To address the communication and computational overhead challenges in traditional FL.
• To enhance privacy preservation and model convergence speed in distributed learning.

Main Methods:

• Integration of quantization compression techniques into a chained FL architecture (Q-Chain FL).
• Efficient compression and transmission of model parameter differences at user nodes.
• Seamless decompression and aggregation of parameters at the server node.

Main Results:

• Q-Chain FL demonstrates low communication and computational overhead.
• The framework achieves fast convergence speed and high security across multiple datasets (MNIST, CIFAR-10, CelebA).
• Reductions in communication overhead of approximately 62.5% (vs. FedAvg) and 44.7% (vs. Chain-PPFL) were observed.

Conclusions:

• Q-Chain FL offers a robust and adaptable solution for federated learning.
• The proposed framework effectively mitigates privacy risks while improving efficiency.
• Results highlight the potential of Q-Chain FL for real-world distributed learning scenarios.