Blockchain-Enabled Trustworthy AI: Decentralized Federated Learning with Proof-of-Learning Consensus

Abstract

We introduce PoL-FL, a blockchain-based FL system where participants submit cryptographic proofs of valid training (Proof-of-Learning) to a smart contract on Ethereum L2 (Optimistic Rollups). PoL-FL combines ZK-STARKs for proving gradient correctness and sharded model aggregation to scale to 10,000 nodes. Tests on IoT sensor data show Byzantine fault tolerance (BFT) under 30% malicious nodes, with 12% lower energy than PoW-based FL.