UNITE: Privacy-Aware Verifiable Quality Assessment via Federated Learning in Blockchain-Empowered Crowdsourcing

As a new type of task execution mode, crowdsourcing makes use of crowd/worker intelligence to collaboratively complete diverse tasks published by task requesters. Quality assessment is an important stage in crowdsourcing as the publicly recruited workers often vary in reliability when performing tasks. Prior works on crowdsourcing quality assessment either ignore the possible privacy disclosure from the task data or are vulnerable to biased evaluation from malicious evaluators. In this paper, we propose a privacy-aware verifiable crowdsourcing quality assessment scheme UNITE against semi-honest and malicious adversaries. UNITE explores federated learning for privacy-aware training of task models, which serves as an indicator of quality assessment. To prevent attackers from deducing the task data from model gradients, we design a secure model update protocol based on differential privacy and perform it with blockchain smart contracts for trustworthy model aggregation. In the presence of malicious requesters providing incorrect assessments, we exploit Pedersen Commitment to generate evidence, which is recorded on-chain with some metadata for public audit. Detailed privacy analysis demonstrates that our differential privacy scheme satisfies (ϵ,δ)-local differential privacy. Finally, we conducted extensive experiments on two real-world datasets and deployed the smart contracts on Hyperledger Fabric to demonstrate good accuracy and both on-chain and off-chain performance.