A polynomial proxy model approach to verifiable decentralized federated learning.

Area of Science:

• Distributed Computing
• Cryptography
• Machine Learning

Background:

• Decentralized Federated Learning (DFL) enhances data privacy but faces challenges in ensuring local computation integrity.
• Current integrity solutions are complex for large models with non-deterministic elements like random dropouts.

Purpose of the Study:

• To introduce ProxyZKP, a novel framework for verifiable computation integrity in DFL.
• To address the limitations of existing methods in handling large-scale, non-deterministic models.

Main Methods:

• ProxyZKP combines Zero-Knowledge Proofs (ZKPs) with polynomial proxy models for secure gradient update verification.
• Local nodes use a private model and a proxy model to mediate training and exchange updates.
• Differential Privacy is integrated to mitigate Gradient Inversion attacks.

Main Results:

• ProxyZKP significantly reduces computational load, with proof generation times 30-50% faster than zk-SNARKs and Bulletproofs.
• The framework demonstrates high parallelization potential via univariate polynomial decomposition.
• Differential Privacy integration maintains competitive model accuracy while enhancing security.

Conclusions:

• ProxyZKP provides a scalable and efficient solution for DFL training integrity.
• The framework is particularly suitable for scenarios requiring frequent updates and robust scalability.
• ProxyZKP effectively balances privacy, integrity, and performance in decentralized learning.