Verifiable Personalized Mutual-Learning based on Blockchain and zk-SNARK

Traditional Federated Learning (FL) trains a single global model by leveraging data from distributed clients, but it struggles in scenarios with heterogeneous and non-i.i.d. client data. Blockchain-based mutual learning has emerged as a promising alternative for personalization, which allows clients to refine local models by exchanging knowledge with peers under the supervision of smart contracts. However, existing solutions often encounter challenges related to communication efficiency, privacy, and verifiability of shared knowledge in trustless or adversarial settings. In this paper, we propose a blockchain-based mutual learning framework that is verifiable and personalized. To ensure the integrity of shared knowledge, the framework incorporates a local endorsement-based mechanism built on a hybrid blockchain architecture and zero-knowledge proofs. For personalization, it allows iterative local model refinement through clustering-based methods for adaptive optimization tailored to each client’s data distribution. Experimental evaluations show that the proposed framework improves performance in heterogeneous data environments while reducing communication overhead compared to existing methods.