PureFL: Channel-partitioned federated learning with cross-channel knowledge distillation on PureChain

Federated learning (FL) improves privacy by keeping data local, but accuracy degrades when clients are partitioned across communication channels and when differential privacy (DP) noise is applied. This paper presents PureFL, a federated learning framework that enhances privacy while addressing scalability issues in blockchain-based FL. It incorporates cross-channel knowledge distillation using a shared public reference dataset with on-chain integrity verification, hybrid on-chain/off-chain aggregation with cryptographic commitments, and formal differential privacy protection with verifiable Rényi DP accounting. PureFL outperforms isolated FL (78.3% vs. 69.4% accuracy on CIFAR-10), reduces on-chain storage by replacing full-update logging with constant-size commitments, and maintains 99.7% synchronization under 20% client unavailability. These improvements are demonstrated across multiple data scenarios.