摘要
Federated Learning (FL) enables collaborative model training across distributed data sources without sharing raw data to enhance privacy, accuracy, and scalability. However, FL remains vulnerable to dishonest aggregators that may tamper with or misuse client model updates, compromising model integrity and privacy. A very natural manner is to authenticate that the aggregated weight is from the individuals' weights. To address this, we propose a verifiable hierarchical peer-to-peer federated learning (P2PFL) framework integrating collaborative zk-SNARKs (proof-size efficient and non-interactive zero knowledge proofs) and multi-party computation. It enables clients to jointly generate a proof that verifies correct aggregation without revealing individual parameters. Also, no single node can access another's plain weights, while allowing multiple clients to efficiently verify the aggregation outcome. Our P2PFL strengthens trust and privacy in federated learning by decentralizing verification and eliminating reliance on a single honest aggregator. © 2026 IEEE.